lerobot/robot-learning-tutorial
508
1@misc{agibot-world-contributorsAgiBotWorldColosseo2025,2 title = {AgiBot World Colosseo: A Large-scale Manipulation Platform for {Scalable} and {Intelligent Embodied Systems}},3 shorttitle = {AgiBot World Colosseo},4 author = {{AgiBot-World-Contributors} and Bu, Qingwen and Cai, Jisong and Chen, Li and Cui, Xiuqi and Ding, Yan and Feng, Siyuan and Gao, Shenyuan and He, Xindong and Hu, Xuan and Huang, Xu and Jiang, Shu and Jiang, Yuxin and Jing, Cheng and Li, Hongyang and Li, Jialu and Liu, Chiming and Liu, Yi and Lu, Yuxiang and Luo, Jianlan and Luo, Ping and Mu, Yao and Niu, Yuehan and Pan, Yixuan and Pang, Jiangmiao and Qiao, Yu and Ren, Guanghui and Ruan, Cheng and Shan, Jiaqi and Shen, Yongjian and Shi, Chengshi and Shi, Mingkang and Shi, Modi and Sima, Chonghao and Song, Jianheng and Wang, Huijie and Wang, Wenhao and Wei, Dafeng and Xie, Chengen and Xu, Guo and Yan, Junchi and Yang, Cunbiao and Yang, Lei and Yang, Shukai and Yao, Maoqing and Zeng, Jia and Zhang, Chi and Zhang, Qinglin and Zhao, Bin and Zhao, Chengyue and Zhao, Jiaqi and Zhu, Jianchao},5 year = {2025},6 month = aug,7 number = {arXiv:2503.06669},8 eprint = {2503.06669},9 primaryclass = {cs},10 publisher = {arXiv},11 doi = {10.48550/arXiv.2503.06669},12 urldate = {2025-08-27},13 abstract = {We explore how scalable robot data can address real-world challenges for generalized robotic manipulation. Introducing AgiBot World, a large-scale platform comprising over 1 million trajectories across 217 tasks in five deployment scenarios, we achieve an order-of-magnitude increase in data scale compared to existing datasets. Accelerated by a standardized collection pipeline with human-in-the-loop verification, AgiBot World guarantees high-quality and diverse data distribution. It is extensible from grippers to dexterous hands and visuo-tactile sensors for fine-grained skill acquisition. Building on top of data, we introduce Genie Operator-1 (GO-1), a novel generalist policy that leverages latent action representations to maximize data utilization, demonstrating predictable performance scaling with increased data volume. Policies pre-trained on our dataset achieve an average performance improvement of 30% over those trained on Open X-Embodiment, both in in-domain and out-of-distribution scenarios. GO-1 exhibits exceptional capability in real-world dexterous and long-horizon tasks, achieving over 60% success rate on complex tasks and outperforming prior RDT approach by 32%. By open-sourcing the dataset, tools, and models, we aim to democratize access to large-scale, high-quality robot data, advancing the pursuit of scalable and general-purpose intelligence.},14 archiveprefix = {arXiv},15 keywords = {Computer Science - Computer Vision and Pattern Recognition,Computer Science - Machine Learning,Computer Science - Robotics},16 file = {/Users/fracapuano/Zotero/storage/TGP4C7GA/AgiBot-World-Contributors et al. - 2025 - AgiBot World Colosseo A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Sys.pdf;/Users/fracapuano/Zotero/storage/IC7BUHWR/2503.html}17}18 19@article{agrawalComputationalSensorimotorLearning,20 title = {Computational {Sensorimotor Learning}},21 author = {Agrawal, Pulkit},22 langid = {english},23 file = {/Users/fracapuano/Zotero/storage/KSDX9GA2/Agrawal - Computational Sensorimotor Learning.pdf}24}25 26@misc{akkayaSolvingRubiksCube2019,27 title = {Solving {Rubik}'s {Cube} with a {Robot Hand}},28 author = {Akkaya, Ilge and Andrychowicz, Marcin and Chociej, Maciek and Litwin, Mateusz and McGrew, Bob and Petron, Arthur and Paino, Alex and Plappert, Matthias and Powell, Glenn and Ribas, Raphael and Schneider, Jonas and Tezak, Nikolas and Tworek, Jerry and Welinder, Peter and Weng, Lilian and Yuan, Qiming and Zaremba, Wojciech and Zhang, Lei},29 year = {2019},30 month = oct,31 number = {arXiv:1910.07113},32 eprint = {1910.07113},33 primaryclass = {cs},34 publisher = {arXiv},35 doi = {10.48550/arXiv.1910.07113},36 urldate = {2025-08-26},37 abstract = {We demonstrate that models trained only in simulation can be used to solve a manipulation problem of unprecedented complexity on a real robot. This is made possible by two key components: a novel algorithm, which we call automatic domain randomization (ADR) and a robot platform built for machine learning. ADR automatically generates a distribution over randomized environments of ever-increasing difficulty. Control policies and vision state estimators trained with ADR exhibit vastly improved sim2real transfer. For control policies, memory-augmented models trained on an ADR-generated distribution of environments show clear signs of emergent meta-learning at test time. The combination of ADR with our custom robot platform allows us to solve a Rubik's cube with a humanoid robot hand, which involves both control and state estimation problems. Videos summarizing our results are available: https://openai.com/blog/solving-rubiks-cube/},38 archiveprefix = {arXiv},39 keywords = {Computer Science - Artificial Intelligence,Computer Science - Computer Vision and Pattern Recognition,Computer Science - Machine Learning,Computer Science - Robotics,Statistics - Machine Learning},40 file = {/Users/fracapuano/Zotero/storage/5HNZLG9D/OpenAI et al. - 2019 - Solving Rubik's Cube with a Robot Hand.pdf;/Users/fracapuano/Zotero/storage/WSM7BJ4I/1910.html}41}42 43@misc{alayracFlamingoVisualLanguage2022,44 title = {Flamingo: A {Visual Language Model} for Few-Shot Learning},45 shorttitle = {Flamingo},46 author = {Alayrac, Jean-Baptiste and Donahue, Jeff and Luc, Pauline and Miech, Antoine and Barr, Iain and Hasson, Yana and Lenc, Karel and Mensch, Arthur and Millican, Katie and Reynolds, Malcolm and Ring, Roman and Rutherford, Eliza and Cabi, Serkan and Han, Tengda and Gong, Zhitao and Samangooei, Sina and Monteiro, Marianne and Menick, Jacob and Borgeaud, Sebastian and Brock, Andrew and Nematzadeh, Aida and Sharifzadeh, Sahand and Binkowski, Mikolaj and Barreira, Ricardo and Vinyals, Oriol and Zisserman, Andrew and Simonyan, Karen},47 year = {2022},48 month = nov,49 number = {arXiv:2204.14198},50 eprint = {2204.14198},51 primaryclass = {cs},52 publisher = {arXiv},53 doi = {10.48550/arXiv.2204.14198},54 urldate = {2025-08-27},55 abstract = {Building models that can be rapidly adapted to novel tasks using only a handful of annotated examples is an open challenge for multimodal machine learning research. We introduce Flamingo, a family of Visual Language Models (VLM) with this ability. We propose key architectural innovations to: (i) bridge powerful pretrained vision-only and language-only models, (ii) handle sequences of arbitrarily interleaved visual and textual data, and (iii) seamlessly ingest images or videos as inputs. Thanks to their flexibility, Flamingo models can be trained on large-scale multimodal web corpora containing arbitrarily interleaved text and images, which is key to endow them with in-context few-shot learning capabilities. We perform a thorough evaluation of our models, exploring and measuring their ability to rapidly adapt to a variety of image and video tasks. These include open-ended tasks such as visual question-answering, where the model is prompted with a question which it has to answer; captioning tasks, which evaluate the ability to describe a scene or an event; and close-ended tasks such as multiple-choice visual question-answering. For tasks lying anywhere on this spectrum, a single Flamingo model can achieve a new state of the art with few-shot learning, simply by prompting the model with task-specific examples. On numerous benchmarks, Flamingo outperforms models fine-tuned on thousands of times more task-specific data.},56 archiveprefix = {arXiv},57 keywords = {Computer Science - Artificial Intelligence,Computer Science - Computer Vision and Pattern Recognition,Computer Science - Machine Learning},58 file = {/Users/fracapuano/Zotero/storage/QZ69HN5K/Alayrac et al. - 2022 - Flamingo a Visual Language Model for Few-Shot Learning.pdf;/Users/fracapuano/Zotero/storage/JMAD5HJY/2204.html}59}60 61@article{aldacoALOHA2Enhanced,62 title = {ALOHA 2: An Enhanced Low-Cost Hardware for {Bimanual Teleoperation}},63 author = {Aldaco, Jorge and Armstrong, Travis and Baruch, Robert and Bingham, Jeff and Chan, Sanky and Dwibedi, Debidatta and Finn, Chelsea and Florence, Pete and Goodrich, Spencer and Gramlich, Wayne and Herzog, Alexander and Hoech, Jonathan and Nguyen, Thinh and Storz, Ian and Tabanpour, Baruch and Tompson, Jonathan and Wahid, Ayzaan and Wahrburg, Ted and Xu, Sichun and Yaroshenko, Sergey and Zhao, Tony Z},64 langid = {english},65 file = {/Users/fracapuano/Zotero/storage/LDEJG62Q/Aldaco et al. - ALOHA 2 An Enhanced Low-Cost Hardware for Bimanual Teleoperation.pdf}66}67 68@article{alizadehComprehensiveSurveySpace2024,69 title = {A Comprehensive Survey of Space Robotic Manipulators for On-Orbit Servicing},70 author = {Alizadeh, Mohammad and Zhu, Zheng H.},71 year = {2024},72 month = oct,73 journal = {Frontiers in Robotics and AI},74 volume = {11},75 publisher = {Frontiers},76 issn = {2296-9144},77 doi = {10.3389/frobt.2024.1470950},78 urldate = {2025-08-26},79 abstract = {On-Orbit Servicing (OOS) robots are transforming space exploration by enabling vital maintenance and repair of spacecraft directly in space. However, achieving precise and safe manipulation in microgravity necessitates overcoming significant challenges. This survey delves into four crucial areas essential for successful OOS manipulation: object state estimation, motion planning, and feedback control. Techniques from traditional vision to advanced X-ray and neural network methods are explored for object state estimation. Strategies for fuel-optimized trajectories, docking maneuvers, and collision avoidance are examined in motion planning. The survey also explores control methods for various scenarios, including cooperative manipulation and handling uncertainties, in feedback control. Additionally, this survey examines how Machine learning techniques can further propel OOS robots towards more complex and delicate tasks in space.},80 langid = {english},81 keywords = {control,machine learning,motion planning,on-orbit servicing,pose estimation,robotic manipulator,space robots},82 file = {/Users/fracapuano/Zotero/storage/VA36KZYY/Alizadeh and Zhu - 2024 - A comprehensive survey of space robotic manipulators for on-orbit servicing.pdf}83}84 85@misc{allalSmolLM2WhenSmol2025,86 title = {SmolLM2: {When Smol Goes Big} -- Data-Centric Training of a {Small Language Model}},87 shorttitle = {SmolLM2},88 author = {Allal, Loubna Ben and Lozhkov, Anton and Bakouch, Elie and Bl{\'a}zquez, Gabriel Mart{\'i}n and Penedo, Guilherme and Tunstall, Lewis and Marafioti, Andr{\'e}s and Kydl{\'i}{\v c}ek, Hynek and Lajar{\'i}n, Agust{\'i}n Piqueres and Srivastav, Vaibhav and Lochner, Joshua and Fahlgren, Caleb and Nguyen, Xuan-Son and Fourrier, Cl{\'e}mentine and Burtenshaw, Ben and Larcher, Hugo and Zhao, Haojun and Zakka, Cyril and Morlon, Mathieu and Raffel, Colin and von Werra, Leandro and Wolf, Thomas},89 year = {2025},90 month = feb,91 number = {arXiv:2502.02737},92 eprint = {2502.02737},93 primaryclass = {cs},94 publisher = {arXiv},95 doi = {10.48550/arXiv.2502.02737},96 urldate = {2025-09-09},97 abstract = {While large language models have facilitated breakthroughs in many applications of artificial intelligence, their inherent largeness makes them computationally expensive and challenging to deploy in resource-constrained settings. In this paper, we document the development of SmolLM2, a state-of-the-art "small" (1.7 billion parameter) language model (LM). To attain strong performance, we overtrain SmolLM2 on {\textasciitilde}11 trillion tokens of data using a multi-stage training process that mixes web text with specialized math, code, and instruction-following data. We additionally introduce new specialized datasets (FineMath, Stack-Edu, and SmolTalk) at stages where we found existing datasets to be problematically small or low-quality. To inform our design decisions, we perform both small-scale ablations as well as a manual refinement process that updates the dataset mixing rates at each stage based on the performance at the previous stage. Ultimately, we demonstrate that SmolLM2 outperforms other recent small LMs including Qwen2.5-1.5B and Llama3.2-1B. To facilitate future research on LM development as well as applications of small LMs, we release both SmolLM2 as well as all of the datasets we prepared in the course of this project.},98 archiveprefix = {arXiv},99 keywords = {Computer Science - Computation and Language},100 file = {/Users/fracapuano/Zotero/storage/I7XDMSV7/Allal et al. - 2025 - SmolLM2 When Smol Goes Big -- Data-Centric Training of a Small Language Model.pdf;/Users/fracapuano/Zotero/storage/6MLZI84T/2502.html}101}102 103@misc{antonovaReinforcementLearningPivoting2017,104 title = {Reinforcement {Learning} for {Pivoting Task}},105 author = {Antonova, Rika and Cruciani, Silvia and Smith, Christian and Kragic, Danica},106 year = {2017},107 month = mar,108 number = {arXiv:1703.00472},109 eprint = {1703.00472},110 primaryclass = {cs},111 publisher = {arXiv},112 doi = {10.48550/arXiv.1703.00472},113 urldate = {2025-08-25},114 abstract = {In this work we propose an approach to learn a robust policy for solving the pivoting task. Recently, several model-free continuous control algorithms were shown to learn successful policies without prior knowledge of the dynamics of the task. However, obtaining successful policies required thousands to millions of training episodes, limiting the applicability of these approaches to real hardware. We developed a training procedure that allows us to use a simple custom simulator to learn policies robust to the mismatch of simulation vs robot. In our experiments, we demonstrate that the policy learned in the simulator is able to pivot the object to the desired target angle on the real robot. We also show generalization to an object with different inertia, shape, mass and friction properties than those used during training. This result is a step towards making model-free reinforcement learning available for solving robotics tasks via pre-training in simulators that offer only an imprecise match to the real-world dynamics.},115 archiveprefix = {arXiv},116 keywords = {Computer Science - Machine Learning,Computer Science - Robotics},117 file = {/Users/fracapuano/Zotero/storage/WRZCHVGB/Antonova et al. - 2017 - Reinforcement Learning for Pivoting Task.pdf;/Users/fracapuano/Zotero/storage/WJEJ2VGU/1703.html}118}119 120@article{aractingiControllingSolo12Quadruped2023,121 title = {Controlling the Solo12 Quadruped Robot with Deep Reinforcement Learning},122 author = {Aractingi, Michel and L{\'e}ziart, Pierre-Alexandre and Flayols, Thomas and Perez, Julien and Silander, Tomi and Sou{\`e}res, Philippe},123 year = {2023},124 month = jul,125 journal = {Scientific Reports},126 volume = {13},127 number = {1},128 pages = {11945},129 publisher = {Nature Publishing Group},130 issn = {2045-2322},131 doi = {10.1038/s41598-023-38259-7},132 urldate = {2025-08-27},133 abstract = {Quadruped robots require robust and general locomotion skills to exploit their mobility potential in complex and challenging environments. In this work, we present an implementation of a robust end-to-end learning-based controller on the Solo12 quadruped. Our method is based on deep reinforcement learning of joint impedance references. The resulting control policies follow a commanded velocity reference while being efficient in its energy consumption and easy to deploy. We detail the learning procedure and method for transfer on the real robot. We show elaborate experiments. Finally, we present experimental results of the learned locomotion on various grounds indoors and outdoors. These results show that the Solo12 robot is a suitable open-source platform for research combining learning and control because of the easiness in transferring and deploying learned controllers.},134 copyright = {2023 The Author(s)},135 langid = {english},136 keywords = {Computer science,Information technology},137 file = {/Users/fracapuano/Zotero/storage/84ZFT7RP/Aractingi et al. - 2023 - Controlling the Solo12 quadruped robot with deep reinforcement learning.pdf}138}139 140@misc{bai2025qwen25vl,141 title = {Qwen2.5-{VL} Technical Report},142 author = {Bai, Shuai and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and Song, Sibo and Dang, Kai and Wang, Peng and Wang, Shijie and Tang, Jun and Zhong, Humen and Zhu, Yuanzhi and Yang, Mingkun and Li, Zhaohai and Wan, Jianqiang and Wang, Pengfei and Ding, Wei and Fu, Zheren and Xu, Yiheng and Ye, Jiabo and Zhang, Xi and Xie, Tianbao and Cheng, Zesen and Zhang, Hang and Yang, Zhibo and Xu, Haiyang and Lin, Junyang},143 year = {2025},144 eprint = {2502.13923},145 primaryclass = {cs.CV},146 archiveprefix = {arXiv}147}148 149@misc{ballEfficientOnlineReinforcement2023,150 title = {Efficient {Online Reinforcement Learning} with {Offline Data}},151 author = {Ball, Philip J. and Smith, Laura and Kostrikov, Ilya and Levine, Sergey},152 year = {2023},153 month = may,154 number = {arXiv:2302.02948},155 eprint = {2302.02948},156 primaryclass = {cs},157 publisher = {arXiv},158 doi = {10.48550/arXiv.2302.02948},159 urldate = {2025-08-30},160 abstract = {Sample efficiency and exploration remain major challenges in online reinforcement learning (RL). A powerful approach that can be applied to address these issues is the inclusion of offline data, such as prior trajectories from a human expert or a sub-optimal exploration policy. Previous methods have relied on extensive modifications and additional complexity to ensure the effective use of this data. Instead, we ask: can we simply apply existing off-policy methods to leverage offline data when learning online? In this work, we demonstrate that the answer is yes; however, a set of minimal but important changes to existing off-policy RL algorithms are required to achieve reliable performance. We extensively ablate these design choices, demonstrating the key factors that most affect performance, and arrive at a set of recommendations that practitioners can readily apply, whether their data comprise a small number of expert demonstrations or large volumes of sub-optimal trajectories. We see that correct application of these simple recommendations can provide a ${\textbackslash}mathbf\{2.5{\textbackslash}times\}$ improvement over existing approaches across a diverse set of competitive benchmarks, with no additional computational overhead. We have released our code at https://github.com/ikostrikov/rlpd.},161 archiveprefix = {arXiv},162 keywords = {Computer Science - Artificial Intelligence,Computer Science - Machine Learning},163 file = {/Users/fracapuano/Zotero/storage/MUKA5D2V/Ball et al. - 2023 - Efficient Online Reinforcement Learning with Offline Data.pdf;/Users/fracapuano/Zotero/storage/IKURHC3D/2302.html}164}165 166@misc{bekrisStateRobotMotion2024,167 title = {The {State} of {Robot Motion Generation}},168 author = {Bekris, Kostas E. and Doerr, Joe and Meng, Patrick and Tangirala, Sumanth},169 year = {2024},170 month = oct,171 number = {arXiv:2410.12172},172 eprint = {2410.12172},173 primaryclass = {cs},174 publisher = {arXiv},175 doi = {10.48550/arXiv.2410.12172},176 urldate = {2025-08-26},177 abstract = {This paper reviews the large spectrum of methods for generating robot motion proposed over the 50 years of robotics research culminating in recent developments. It crosses the boundaries of methodologies, typically not surveyed together, from those that operate over explicit models to those that learn implicit ones. The paper discusses the current state-of-the-art as well as properties of varying methodologies, highlighting opportunities for integration.},178 archiveprefix = {arXiv},179 keywords = {Computer Science - Artificial Intelligence,Computer Science - Machine Learning,Computer Science - Robotics},180 file = {/Users/fracapuano/Zotero/storage/DMJJZFDZ/Bekris et al. - 2024 - The State of Robot Motion Generation.pdf;/Users/fracapuano/Zotero/storage/TL42IRAN/2410.html}181}182 183@article{bellemareAutonomousNavigationStratospheric2020,184 title = {Autonomous Navigation of Stratospheric Balloons Using Reinforcement Learning},185 author = {Bellemare, Marc G. and Candido, Salvatore and Castro, Pablo Samuel and Gong, Jun and Machado, Marlos C. and Moitra, Subhodeep and Ponda, Sameera S. and Wang, Ziyu},186 year = {2020},187 month = dec,188 journal = {Nature},189 volume = {588},190 number = {7836},191 pages = {77--82},192 publisher = {Nature Publishing Group},193 issn = {1476-4687},194 doi = {10.1038/s41586-020-2939-8},195 urldate = {2025-08-31},196 abstract = {Efficiently navigating a superpressure balloon in the stratosphere1 requires the integration of a multitude of cues, such as wind speed and solar elevation, and the process is complicated by forecast errors and sparse wind measurements. Coupled with the need to make decisions in real time, these factors rule out the use of conventional control techniques2,3. Here we describe the use of reinforcement learning4,5 to create a high-performing flight controller. Our algorithm uses data augmentation6,7 and a self-correcting design to overcome the key technical challenge of reinforcement learning from imperfect data, which has proved to be a major obstacle to its application to physical systems8. We deployed our controller to station Loon superpressure balloons at multiple locations across the globe, including a 39-day controlled experiment over the Pacific Ocean. Analyses show that the controller outperforms Loon's previous algorithm and is robust to the natural diversity in stratospheric winds. These results demonstrate that reinforcement learning is an effective solution to real-world autonomous control problems in which neither conventional methods nor human intervention suffice, offering clues about what may be needed to create artificially intelligent agents that continuously interact with real, dynamic environments.},197 copyright = {2020 The Author(s), under exclusive licence to Springer Nature Limited},198 langid = {english},199 keywords = {Aerospace engineering,Computer science}200}201 202@article{bellmanMarkovianDecisionProcess1957,203 title = {A {Markovian Decision Process}},204 author = {Bellman, Richard},205 year = {1957},206 journal = {Journal of Mathematics and Mechanics},207 volume = {6},208 number = {5},209 eprint = {24900506},210 eprinttype = {jstor},211 pages = {679--684},212 publisher = {Indiana University Mathematics Department},213 issn = {0095-9057},214 urldate = {2025-08-30}215}216 217@misc{beyerPaliGemmaVersatile3B2024,218 title = {PaliGemma: {A} Versatile 3B VLM for Transfer},219 shorttitle = {PaliGemma},220 author = {Beyer, Lucas and Steiner, Andreas and Pinto, Andr{\'e} Susano and Kolesnikov, Alexander and Wang, Xiao and Salz, Daniel and Neumann, Maxim and Alabdulmohsin, Ibrahim and Tschannen, Michael and Bugliarello, Emanuele and Unterthiner, Thomas and Keysers, Daniel and Koppula, Skanda and Liu, Fangyu and Grycner, Adam and Gritsenko, Alexey and Houlsby, Neil and Kumar, Manoj and Rong, Keran and Eisenschlos, Julian and Kabra, Rishabh and Bauer, Matthias and Bo{\v s}njak, Matko and Chen, Xi and Minderer, Matthias and Voigtlaender, Paul and Bica, Ioana and Balazevic, Ivana and Puigcerver, Joan and Papalampidi, Pinelopi and Henaff, Olivier and Xiong, Xi and Soricut, Radu and Harmsen, Jeremiah and Zhai, Xiaohua},221 year = {2024},222 month = oct,223 number = {arXiv:2407.07726},224 eprint = {2407.07726},225 primaryclass = {cs},226 publisher = {arXiv},227 doi = {10.48550/arXiv.2407.07726},228 urldate = {2025-09-08},229 abstract = {PaliGemma is an open Vision-Language Model (VLM) that is based on the SigLIP-So400m vision encoder and the Gemma-2B language model. It is trained to be a versatile and broadly knowledgeable base model that is effective to transfer. It achieves strong performance on a wide variety of open-world tasks. We evaluate PaliGemma on almost 40 diverse tasks including standard VLM benchmarks, but also more specialized tasks such as remote-sensing and segmentation.},230 archiveprefix = {arXiv},231 keywords = {Computer Science - Artificial Intelligence,Computer Science - Computation and Language,Computer Science - Computer Vision and Pattern Recognition,Computer Science - Machine Learning},232 file = {/Users/fracapuano/Zotero/storage/IPDYNWC4/Beyer et al. - 2024 - PaliGemma A versatile 3B VLM for transfer.pdf;/Users/fracapuano/Zotero/storage/R7UVD9WC/2407.html}233}234 235@misc{bjorckGR00TN1Open2025,236 title = {GR00T N1: {An Open Foundation Model} for {Generalist Humanoid Robots}},237 shorttitle = {GR00T N1},238 author = {Bjorck, Johan and Casta{\~n}eda, Fernando and Cherniadev, Nikita and Da, Xingye and Ding, Runyu and Fan, Linxi "Jim" and Fang, Yu and Fox, Dieter and Hu, Fengyuan and Huang, Spencer and Jang, Joel and Jiang, Zhenyu and Kautz, Jan and Kundalia, Kaushil and Lao, Lawrence and Li, Zhiqi and Lin, Zongyu and Lin, Kevin and Liu, Guilin and Llontop, Edith and Magne, Loic and Mandlekar, Ajay and Narayan, Avnish and Nasiriany, Soroush and Reed, Scott and Tan, You Liang and Wang, Guanzhi and Wang, Zu and Wang, Jing and Wang, Qi and Xiang, Jiannan and Xie, Yuqi and Xu, Yinzhen and Xu, Zhenjia and Ye, Seonghyeon and Yu, Zhiding and Zhang, Ao and Zhang, Hao and Zhao, Yizhou and Zheng, Ruijie and Zhu, Yuke},239 year = {2025},240 month = mar,241 number = {arXiv:2503.14734},242 eprint = {2503.14734},243 primaryclass = {cs},244 publisher = {arXiv},245 doi = {10.48550/arXiv.2503.14734},246 urldate = {2025-08-26},247 abstract = {General-purpose robots need a versatile body and an intelligent mind. Recent advancements in humanoid robots have shown great promise as a hardware platform for building generalist autonomy in the human world. A robot foundation model, trained on massive and diverse data sources, is essential for enabling the robots to reason about novel situations, robustly handle real-world variability, and rapidly learn new tasks. To this end, we introduce GR00T N1, an open foundation model for humanoid robots. GR00T N1 is a Vision-Language-Action (VLA) model with a dual-system architecture. The vision-language module (System 2) interprets the environment through vision and language instructions. The subsequent diffusion transformer module (System 1) generates fluid motor actions in real time. Both modules are tightly coupled and jointly trained end-to-end. We train GR00T N1 with a heterogeneous mixture of real-robot trajectories, human videos, and synthetically generated datasets. We show that our generalist robot model GR00T N1 outperforms the state-of-the-art imitation learning baselines on standard simulation benchmarks across multiple robot embodiments. Furthermore, we deploy our model on the Fourier GR-1 humanoid robot for language-conditioned bimanual manipulation tasks, achieving strong performance with high data efficiency.},248 archiveprefix = {arXiv},249 keywords = {Computer Science - Artificial Intelligence,Computer Science - Machine Learning,Computer Science - Robotics},250 file = {/Users/fracapuano/Zotero/storage/BDNSKFA6/NVIDIA et al. - 2025 - GR00T N1 An Open Foundation Model for Generalist Humanoid Robots.pdf;/Users/fracapuano/Zotero/storage/FENU9PQR/2503.html}251}252 253@misc{black$p_0$VisionLanguageActionFlow2024,254 title = {$\pi_0$: A Vision-Language-Action Flow Model for {General Robot Control}},255 shorttitle = {$\pi_0$},256 author = {Black, Kevin and Brown, Noah and Driess, Danny and Esmail, Adnan and Equi, Michael and Finn, Chelsea and Fusai, Niccolo and Groom, Lachy and Hausman, Karol and Ichter, Brian and Jakubczak, Szymon and Jones, Tim and Ke, Liyiming and Levine, Sergey and {Li-Bell}, Adrian and Mothukuri, Mohith and Nair, Suraj and Pertsch, Karl and Shi, Lucy Xiaoyang and Tanner, James and Vuong, Quan and Walling, Anna and Wang, Haohuan and Zhilinsky, Ury},257 year = {2024},258 month = oct,259 number = {arXiv:2410.24164},260 eprint = {2410.24164},261 primaryclass = {cs},262 publisher = {arXiv},263 doi = {10.48550/arXiv.2410.24164},264 urldate = {2025-08-28},265 abstract = {Robot learning holds tremendous promise to unlock the full potential of flexible, general, and dexterous robot systems, as well as to address some of the deepest questions in artificial intelligence. However, bringing robot learning to the level of generality required for effective real-world systems faces major obstacles in terms of data, generalization, and robustness. In this paper, we discuss how generalist robot policies (i.e., robot foundation models) can address these challenges, and how we can design effective generalist robot policies for complex and highly dexterous tasks. We propose a novel flow matching architecture built on top of a pre-trained vision-language model (VLM) to inherit Internet-scale semantic knowledge. We then discuss how this model can be trained on a large and diverse dataset from multiple dexterous robot platforms, including single-arm robots, dual-arm robots, and mobile manipulators. We evaluate our model in terms of its ability to perform tasks in zero shot after pre-training, follow language instructions from people and from a high-level VLM policy, and its ability to acquire new skills via fine-tuning. Our results cover a wide variety of tasks, such as laundry folding, table cleaning, and assembling boxes.},266 archiveprefix = {arXiv},267 keywords = {Computer Science - Machine Learning,Computer Science - Robotics},268 file = {/Users/fracapuano/Zotero/storage/GUEM37NZ/Black et al. - 2024 - $π_0$ A Vision-Language-Action Flow Model for General Robot Control.pdf;/Users/fracapuano/Zotero/storage/FHYXZWF8/2410.html}269}270 271@inproceedings{BLIP-2,272 title = {BLIP-2: Bootstrapping Language-Image Pre-Training with Frozen Image Encoders and Large Language Models},273 booktitle = {Proceedings of the 40th International Conference on Machine Learning},274 author = {Li, Junnan and Li, Dongxu and Savarese, Silvio and Hoi, Steven},275 year = {2023},276 series = {ICML'23},277 publisher = {JMLR.org},278 address = {, Honolulu, Hawaii, USA,},279 abstract = {The cost of vision-and-language pre-training has become increasingly prohibitive due to end-to-end training of large-scale models. This paper proposes BLIP-2, a generic and efficient pretraining strategy that bootstraps vision-language pre-training from off-the-shelf frozen pretrained image encoders and frozen large language models. BLIP-2 bridges the modality gap with a lightweight Querying Transformer, which is pretrained in two stages. The first stage bootstraps vision-language representation learning from a frozen image encoder. The second stage bootstraps vision-to-language generative learning from a frozen language model. BLIP-2 achieves state-of-the-art performance on various vision-language tasks, despite having significantly fewer trainable parameters than existing methods. For example, our model outperforms Flamingo80B by 8.7% on zero-shot VQAv2 with 54x fewer trainable parameters. We also demonstrate the model's capabilities of zero-shot image-to-text generation that can follow natural language instructions.},280 articleno = {814}281}282 283@misc{brohanRT1RoboticsTransformer2023,284 title = {RT-1: {Robotics Transformer} for Real-World Control at {Scale}},285 shorttitle = {RT-1},286 author = {Brohan, Anthony and Brown, Noah and Carbajal, Justice and Chebotar, Yevgen and Dabis, Joseph and Finn, Chelsea and Gopalakrishnan, Keerthana and Hausman, Karol and Herzog, Alex and Hsu, Jasmine and Ibarz, Julian and Ichter, Brian and Irpan, Alex and Jackson, Tomas and Jesmonth, Sally and Joshi, Nikhil J. and Julian, Ryan and Kalashnikov, Dmitry and Kuang, Yuheng and Leal, Isabel and Lee, Kuang-Huei and Levine, Sergey and Lu, Yao and Malla, Utsav and Manjunath, Deeksha and Mordatch, Igor and Nachum, Ofir and Parada, Carolina and Peralta, Jodilyn and Perez, Emily and Pertsch, Karl and Quiambao, Jornell and Rao, Kanishka and Ryoo, Michael and Salazar, Grecia and Sanketi, Pannag and Sayed, Kevin and Singh, Jaspiar and Sontakke, Sumedh and Stone, Austin and Tan, Clayton and Tran, Huong and Vanhoucke, Vincent and Vega, Steve and Vuong, Quan and Xia, Fei and Xiao, Ted and Xu, Peng and Xu, Sichun and Yu, Tianhe and Zitkovich, Brianna},287 year = {2023},288 month = aug,289 number = {arXiv:2212.06817},290 eprint = {2212.06817},291 primaryclass = {cs},292 publisher = {arXiv},293 doi = {10.48550/arXiv.2212.06817},294 urldate = {2025-09-07},295 abstract = {By transferring knowledge from large, diverse, task-agnostic datasets, modern machine learning models can solve specific downstream tasks either zero-shot or with small task-specific datasets to a high level of performance. While this capability has been demonstrated in other fields such as computer vision, natural language processing or speech recognition, it remains to be shown in robotics, where the generalization capabilities of the models are particularly critical due to the difficulty of collecting real-world robotic data. We argue that one of the keys to the success of such general robotic models lies with open-ended task-agnostic training, combined with high-capacity architectures that can absorb all of the diverse, robotic data. In this paper, we present a model class, dubbed Robotics Transformer, that exhibits promising scalable model properties. We verify our conclusions in a study of different model classes and their ability to generalize as a function of the data size, model size, and data diversity based on a large-scale data collection on real robots performing real-world tasks. The project's website and videos can be found at robotics-transformer1.github.io},296 archiveprefix = {arXiv},297 keywords = {Computer Science - Artificial Intelligence,Computer Science - Computation and Language,Computer Science - Computer Vision and Pattern Recognition,Computer Science - Machine Learning,Computer Science - Robotics},298 file = {/Users/fracapuano/Zotero/storage/TTBN3M5Y/Brohan et al. - 2023 - RT-1 Robotics Transformer for Real-World Control at Scale.pdf;/Users/fracapuano/Zotero/storage/DK3D593W/2212.html}299}300 301@misc{brohanRT2VisionLanguageActionModels2023,302 title = {RT-2: Vision-Language-Action Models Transfer Web Knowledge to {Robotic Control}},303 shorttitle = {RT-2},304 author = {Brohan, Anthony and Brown, Noah and Carbajal, Justice and Chebotar, Yevgen and Chen, Xi and Choromanski, Krzysztof and Ding, Tianli and Driess, Danny and Dubey, Avinava and Finn, Chelsea and Florence, Pete and Fu, Chuyuan and Arenas, Montse Gonzalez and Gopalakrishnan, Keerthana and Han, Kehang and Hausman, Karol and Herzog, Alexander and Hsu, Jasmine and Ichter, Brian and Irpan, Alex and Joshi, Nikhil and Julian, Ryan and Kalashnikov, Dmitry and Kuang, Yuheng and Leal, Isabel and Lee, Lisa and Lee, Tsang-Wei Edward and Levine, Sergey and Lu, Yao and Michalewski, Henryk and Mordatch, Igor and Pertsch, Karl and Rao, Kanishka and Reymann, Krista and Ryoo, Michael and Salazar, Grecia and Sanketi, Pannag and Sermanet, Pierre and Singh, Jaspiar and Singh, Anikait and Soricut, Radu and Tran, Huong and Vanhoucke, Vincent and Vuong, Quan and Wahid, Ayzaan and Welker, Stefan and Wohlhart, Paul and Wu, Jialin and Xia, Fei and Xiao, Ted and Xu, Peng and Xu, Sichun and Yu, Tianhe and Zitkovich, Brianna},305 year = {2023},306 month = jul,307 number = {arXiv:2307.15818},308 eprint = {2307.15818},309 primaryclass = {cs},310 publisher = {arXiv},311 doi = {10.48550/arXiv.2307.15818},312 urldate = {2025-09-07},313 abstract = {We study how vision-language models trained on Internet-scale data can be incorporated directly into end-to-end robotic control to boost generalization and enable emergent semantic reasoning. Our goal is to enable a single end-to-end trained model to both learn to map robot observations to actions and enjoy the benefits of large-scale pretraining on language and vision-language data from the web. To this end, we propose to co-fine-tune state-of-the-art vision-language models on both robotic trajectory data and Internet-scale vision-language tasks, such as visual question answering. In contrast to other approaches, we propose a simple, general recipe to achieve this goal: in order to fit both natural language responses and robotic actions into the same format, we express the actions as text tokens and incorporate them directly into the training set of the model in the same way as natural language tokens. We refer to such category of models as vision-language-action models (VLA) and instantiate an example of such a model, which we call RT-2. Our extensive evaluation (6k evaluation trials) shows that our approach leads to performant robotic policies and enables RT-2 to obtain a range of emergent capabilities from Internet-scale training. This includes significantly improved generalization to novel objects, the ability to interpret commands not present in the robot training data (such as placing an object onto a particular number or icon), and the ability to perform rudimentary reasoning in response to user commands (such as picking up the smallest or largest object, or the one closest to another object). We further show that incorporating chain of thought reasoning allows RT-2 to perform multi-stage semantic reasoning, for example figuring out which object to pick up for use as an improvised hammer (a rock), or which type of drink is best suited for someone who is tired (an energy drink).},314 archiveprefix = {arXiv},315 keywords = {Computer Science - Computation and Language,Computer Science - Computer Vision and Pattern Recognition,Computer Science - Machine Learning,Computer Science - Robotics},316 file = {/Users/fracapuano/Zotero/storage/CZHMNYPG/Brohan et al. - 2023 - RT-2 Vision-Language-Action Models Transfer Web Knowledge to Robotic Control.pdf;/Users/fracapuano/Zotero/storage/WN2E7AZH/2307.html}317}318 319@misc{brownLanguageModelsAre2020,320 title = {Language {Models} Are Few-Shot Learners},321 author = {Brown, Tom B. and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and Agarwal, Sandhini and {Herbert-Voss}, Ariel and Krueger, Gretchen and Henighan, Tom and Child, Rewon and Ramesh, Aditya and Ziegler, Daniel M. and Wu, Jeffrey and Winter, Clemens and Hesse, Christopher and Chen, Mark and Sigler, Eric and Litwin, Mateusz and Gray, Scott and Chess, Benjamin and Clark, Jack and Berner, Christopher and McCandlish, Sam and Radford, Alec and Sutskever, Ilya and Amodei, Dario},322 year = {2020},323 month = jul,324 number = {arXiv:2005.14165},325 eprint = {2005.14165},326 primaryclass = {cs},327 publisher = {arXiv},328 doi = {10.48550/arXiv.2005.14165},329 urldate = {2025-08-28},330 abstract = {Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task. While typically task-agnostic in architecture, this method still requires task-specific fine-tuning datasets of thousands or tens of thousands of examples. By contrast, humans can generally perform a new language task from only a few examples or from simple instructions - something which current NLP systems still largely struggle to do. Here we show that scaling up language models greatly improves task-agnostic, few-shot performance, sometimes even reaching competitiveness with prior state-of-the-art fine-tuning approaches. Specifically, we train GPT-3, an autoregressive language model with 175 billion parameters, 10x more than any previous non-sparse language model, and test its performance in the few-shot setting. For all tasks, GPT-3 is applied without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction with the model. GPT-3 achieves strong performance on many NLP datasets, including translation, question-answering, and cloze tasks, as well as several tasks that require on-the-fly reasoning or domain adaptation, such as unscrambling words, using a novel word in a sentence, or performing 3-digit arithmetic. At the same time, we also identify some datasets where GPT-3's few-shot learning still struggles, as well as some datasets where GPT-3 faces methodological issues related to training on large web corpora. Finally, we find that GPT-3 can generate samples of news articles which human evaluators have difficulty distinguishing from articles written by humans. We discuss broader societal impacts of this finding and of GPT-3 in general.},331 archiveprefix = {arXiv},332 keywords = {Computer Science - Computation and Language},333 file = {/Users/fracapuano/Zotero/storage/L6J45ZW7/Brown et al. - 2020 - Language Models are Few-Shot Learners.pdf;/Users/fracapuano/Zotero/storage/52DC5AT2/2005.html}334}335 336@article{burridgeSequentialCompositionDynamically1999b,337 title = {Sequential {Composition} of {Dynamically Dexterous Robot Behaviors}},338 author = {Burridge, R. R. and Rizzi, A. A. and Koditschek, D. E.},339 year = {1999},340 month = jun,341 journal = {The International Journal of Robotics Research},342 volume = {18},343 number = {6},344 pages = {534--555},345 issn = {0278-3649, 1741-3176},346 doi = {10.1177/02783649922066385},347 urldate = {2025-08-26},348 abstract = {We report on our efforts to develop a sequential robot controllercomposition technique in the context of dexterous ``batting'' maneuvers. A robot with a flat paddle is required to strike repeatedly at a thrown ball until the ball is brought to rest on the paddle at a specified location. The robot's reachable workspace is blocked by an obstacle that disconnects the free space formed when the ball and paddle remain in contact, forcing the machine to ``let go'' for a time to bring the ball to the desired state. The controller compositions we create guarantee that a ball introduced in the ``safe workspace'' remains there and is ultimately brought to the goal. We report on experimental results from an implementation of these formal composition methods, and present descriptive statistics characterizing the experiments.},349 copyright = {https://journals.sagepub.com/page/policies/text-and-data-mining-license},350 langid = {english},351 file = {/Users/fracapuano/Zotero/storage/TFZQ6EHJ/Burridge et al. - 1999 - Sequential Composition of Dynamically Dexterous Robot Behaviors.pdf}352}353 354@misc{cadene2024lerobot,355 title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},356 author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakirova, Dana and Russi, Martino and Capuano, Francesco and Pascal, Caroline and Choghari, Jade and Moss, Jess and Wolf, Thomas},357 year = {2024}358}359 360@misc{cadeneLeRobotStateoftheartMachine,361 title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in {Pytorch}},362 author = {Cadene, Remi}363}364 365@misc{cadeneLeRobotStateoftheartMachine2024,366 title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in {Pytorch}},367 author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Galloudec, Quentin and Zouitine, Adil and Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakirova, Dana and Russi, Martino and Capuano, Francesco and Pascal, Caroline and Chogari, Jade and Moss, Jess and Wolf, Thomas},368 year = {2024}369}370 371@misc{caronEmergingPropertiesSelfSupervised2021,372 title = {Emerging {Properties} in Self-Supervised Vision Transformers},373 author = {Caron, Mathilde and Touvron, Hugo and Misra, Ishan and J{\'e}gou, Herv{\'e} and Mairal, Julien and Bojanowski, Piotr and Joulin, Armand},374 year = {2021},375 month = may,376 number = {arXiv:2104.14294},377 eprint = {2104.14294},378 primaryclass = {cs},379 publisher = {arXiv},380 doi = {10.48550/arXiv.2104.14294},381 urldate = {2025-09-07},382 abstract = {In this paper, we question if self-supervised learning provides new properties to Vision Transformer (ViT) that stand out compared to convolutional networks (convnets). Beyond the fact that adapting self-supervised methods to this architecture works particularly well, we make the following observations: first, self-supervised ViT features contain explicit information about the semantic segmentation of an image, which does not emerge as clearly with supervised ViTs, nor with convnets. Second, these features are also excellent k-NN classifiers, reaching 78.3% top-1 on ImageNet with a small ViT. Our study also underlines the importance of momentum encoder, multi-crop training, and the use of small patches with ViTs. We implement our findings into a simple self-supervised method, called DINO, which we interpret as a form of self-distillation with no labels. We show the synergy between DINO and ViTs by achieving 80.1% top-1 on ImageNet in linear evaluation with ViT-Base.},383 archiveprefix = {arXiv},384 keywords = {Computer Science - Computer Vision and Pattern Recognition},385 file = {/Users/fracapuano/Zotero/storage/AYIY6DTF/Caron et al. - 2021 - Emerging Properties in Self-Supervised Vision Transformers.pdf;/Users/fracapuano/Zotero/storage/EKA7ZN2P/2104.html}386}387 388@inproceedings{chebotar2019closing,389 title = {Closing the Sim-to-Real Loop: {Adapting} Simulation Randomization with Real World Experience},390 booktitle = {2019 International Conference on Robotics and Automation ({ICRA})},391 author = {Chebotar, Yevgen and Handa, Ankur and Makoviychuk, Viktor and Macklin, Miles and Issac, Jan and Ratliff, Nathan and Fox, Dieter},392 year = {2019},393 pages = {8973--8979},394 publisher = {IEEE}395}396 397@inproceedings{chebotarClosingSimtorealLoop2019,398 title = {Closing the Sim-to-Real Loop: {Adapting} Simulation Randomization with Real World Experience},399 shorttitle = {Closing the Sim-to-Real Loop},400 booktitle = {2019 {International Conference} on {Robotics} and {Automation} ({ICRA})},401 author = {Chebotar, Yevgen and Handa, Ankur and Makoviychuk, Viktor and Macklin, Miles and Issac, Jan and Ratliff, Nathan and Fox, Dieter},402 year = {2019},403 pages = {8973--8979},404 publisher = {IEEE},405 urldate = {2025-08-31}406}407 408@misc{chenPaLIXScalingMultilingual2023,409 title = {PaLI-X: {On Scaling} up a {Multilingual Vision} and {Language Model}},410 shorttitle = {PaLI-X},411 author = {Chen, Xi and Djolonga, Josip and Padlewski, Piotr and Mustafa, Basil and Changpinyo, Soravit and Wu, Jialin and Ruiz, Carlos Riquelme and Goodman, Sebastian and Wang, Xiao and Tay, Yi and Shakeri, Siamak and Dehghani, Mostafa and Salz, Daniel and Lucic, Mario and Tschannen, Michael and Nagrani, Arsha and Hu, Hexiang and Joshi, Mandar and Pang, Bo and Montgomery, Ceslee and Pietrzyk, Paulina and Ritter, Marvin and Piergiovanni, A. J. and Minderer, Matthias and Pavetic, Filip and Waters, Austin and Li, Gang and Alabdulmohsin, Ibrahim and Beyer, Lucas and Amelot, Julien and Lee, Kenton and Steiner, Andreas Peter and Li, Yang and Keysers, Daniel and Arnab, Anurag and Xu, Yuanzhong and Rong, Keran and Kolesnikov, Alexander and Seyedhosseini, Mojtaba and Angelova, Anelia and Zhai, Xiaohua and Houlsby, Neil and Soricut, Radu},412 year = {2023},413 month = may,414 number = {arXiv:2305.18565},415 eprint = {2305.18565},416 primaryclass = {cs},417 publisher = {arXiv},418 doi = {10.48550/arXiv.2305.18565},419 urldate = {2025-09-07},420 abstract = {We present the training recipe and results of scaling up PaLI-X, a multilingual vision and language model, both in terms of size of the components and the breadth of its training task mixture. Our model achieves new levels of performance on a wide-range of varied and complex tasks, including multiple image-based captioning and question-answering tasks, image-based document understanding and few-shot (in-context) learning, as well as object detection, video question answering, and video captioning. PaLI-X advances the state-of-the-art on most vision-and-language benchmarks considered (25+ of them). Finally, we observe emerging capabilities, such as complex counting and multilingual object detection, tasks that are not explicitly in the training mix.},421 archiveprefix = {arXiv},422 keywords = {Computer Science - Computation and Language,Computer Science - Computer Vision and Pattern Recognition,Computer Science - Machine Learning},423 file = {/Users/fracapuano/Zotero/storage/UES2DMFM/Chen et al. - 2023 - PaLI-X On Scaling up a Multilingual Vision and Language Model.pdf;/Users/fracapuano/Zotero/storage/LEGNNSHS/2305.html}424}425 426@misc{chiDiffusionPolicyVisuomotor2024,427 title = {Diffusion {Policy}: {Visuomotor Policy Learning} via {Action Diffusion}},428 shorttitle = {Diffusion {Policy}},429 author = {Chi, Cheng and Xu, Zhenjia and Feng, Siyuan and Cousineau, Eric and Du, Yilun and Burchfiel, Benjamin and Tedrake, Russ and Song, Shuran},430 year = {2024},431 month = mar,432 number = {arXiv:2303.04137},433 eprint = {2303.04137},434 primaryclass = {cs},435 publisher = {arXiv},436 doi = {10.48550/arXiv.2303.04137},437 urldate = {2025-08-28},438 abstract = {This paper introduces Diffusion Policy, a new way of generating robot behavior by representing a robot's visuomotor policy as a conditional denoising diffusion process. We benchmark Diffusion Policy across 12 different tasks from 4 different robot manipulation benchmarks and find that it consistently outperforms existing state-of-the-art robot learning methods with an average improvement of 46.9%. Diffusion Policy learns the gradient of the action-distribution score function and iteratively optimizes with respect to this gradient field during inference via a series of stochastic Langevin dynamics steps. We find that the diffusion formulation yields powerful advantages when used for robot policies, including gracefully handling multimodal action distributions, being suitable for high-dimensional action spaces, and exhibiting impressive training stability. To fully unlock the potential of diffusion models for visuomotor policy learning on physical robots, this paper presents a set of key technical contributions including the incorporation of receding horizon control, visual conditioning, and the time-series diffusion transformer. We hope this work will help motivate a new generation of policy learning techniques that are able to leverage the powerful generative modeling capabilities of diffusion models. Code, data, and training details is publicly available diffusion-policy.cs.columbia.edu},439 archiveprefix = {arXiv},440 keywords = {Computer Science - Robotics},441 file = {/Users/fracapuano/Zotero/storage/7XRY3GJX/Chi et al. - 2024 - Diffusion Policy Visuomotor Policy Learning via Action Diffusion.pdf;/Users/fracapuano/Zotero/storage/BBBPKKMZ/2303.html}442}443 444@misc{collaborationOpenXEmbodimentRobotic2025,445 title = {Open X-Embodiment: {Robotic Learning Datasets} and RT-X Models},446 shorttitle = {Open X-Embodiment},447 author = {Collaboration, Open X.-Embodiment and O'Neill, Abby and Rehman, Abdul and Gupta, Abhinav and Maddukuri, Abhiram and Gupta, Abhishek and Padalkar, Abhishek and Lee, Abraham and Pooley, Acorn and Gupta, Agrim and Mandlekar, Ajay and Jain, Ajinkya and Tung, Albert and Bewley, Alex and Herzog, Alex and Irpan, Alex and Khazatsky, Alexander and Rai, Anant and Gupta, Anchit and Wang, Andrew and Kolobov, Andrey and Singh, Anikait and Garg, Animesh and Kembhavi, Aniruddha and Xie, Annie and Brohan, Anthony and Raffin, Antonin and Sharma, Archit and Yavary, Arefeh and Jain, Arhan and Balakrishna, Ashwin and Wahid, Ayzaan and {Burgess-Limerick}, Ben and Kim, Beomjoon and Sch{\"o}lkopf, Bernhard and Wulfe, Blake and Ichter, Brian and Lu, Cewu and Xu, Charles and Le, Charlotte and Finn, Chelsea and Wang, Chen and Xu, Chenfeng and Chi, Cheng and Huang, Chenguang and Chan, Christine and Agia, Christopher and Pan, Chuer and Fu, Chuyuan and Devin, Coline and Xu, Danfei and Morton, Daniel and Driess, Danny and Chen, Daphne and Pathak, Deepak and Shah, Dhruv and B{\"u}chler, Dieter and Jayaraman, Dinesh and Kalashnikov, Dmitry and Sadigh, Dorsa and Johns, Edward and Foster, Ethan and Liu, Fangchen and Ceola, Federico and Xia, Fei and Zhao, Feiyu and Frujeri, Felipe Vieira and Stulp, Freek and Zhou, Gaoyue and Sukhatme, Gaurav S. and Salhotra, Gautam and Yan, Ge and Feng, Gilbert and Schiavi, Giulio and Berseth, Glen and Kahn, Gregory and Yang, Guangwen and Wang, Guanzhi and Su, Hao and Fang, Hao-Shu and Shi, Haochen and Bao, Henghui and Amor, Heni Ben and Christensen, Henrik I. and Furuta, Hiroki and Bharadhwaj, Homanga and Walke, Homer and Fang, Hongjie and Ha, Huy and Mordatch, Igor and Radosavovic, Ilija and Leal, Isabel and Liang, Jacky and {Abou-Chakra}, Jad and Kim, Jaehyung and Drake, Jaimyn and Peters, Jan and Schneider, Jan and Hsu, Jasmine and Vakil, Jay and Bohg, Jeannette and Bingham, Jeffrey and Wu, Jeffrey and Gao, Jensen and Hu, Jiaheng and Wu, Jiajun and Wu, Jialin and Sun, Jiankai and Luo, Jianlan and Gu, Jiayuan and Tan, Jie and Oh, Jihoon and Wu, Jimmy and Lu, Jingpei and Yang, Jingyun and Malik, Jitendra and Silv{\'e}rio, Jo{\~a}o and Hejna, Joey and Booher, Jonathan and Tompson, Jonathan and Yang, Jonathan and Salvador, Jordi and Lim, Joseph J. and Han, Junhyek and Wang, Kaiyuan and Rao, Kanishka and Pertsch, Karl and Hausman, Karol and Go, Keegan and Gopalakrishnan, Keerthana and Goldberg, Ken and Byrne, Kendra and Oslund, Kenneth and Kawaharazuka, Kento and Black, Kevin and Lin, Kevin and Zhang, Kevin and Ehsani, Kiana and Lekkala, Kiran and Ellis, Kirsty and Rana, Krishan and Srinivasan, Krishnan and Fang, Kuan and Singh, Kunal Pratap and Zeng, Kuo-Hao and Hatch, Kyle and Hsu, Kyle and Itti, Laurent and Chen, Lawrence Yunliang and Pinto, Lerrel and {Fei-Fei}, Li and Tan, Liam and Fan, Linxi "Jim" and Ott, Lionel and Lee, Lisa and Weihs, Luca and Chen, Magnum and Lepert, Marion and Memmel, Marius and Tomizuka, Masayoshi and Itkina, Masha and Castro, Mateo Guaman and Spero, Max and Du, Maximilian and Ahn, Michael and Yip, Michael C. and Zhang, Mingtong and Ding, Mingyu and Heo, Minho and Srirama, Mohan Kumar and Sharma, Mohit and Kim, Moo Jin and Irshad, Muhammad Zubair and Kanazawa, Naoaki and Hansen, Nicklas and Heess, Nicolas and Joshi, Nikhil J. and Suenderhauf, Niko and Liu, Ning and Palo, Norman Di and Shafiullah, Nur Muhammad Mahi and Mees, Oier and Kroemer, Oliver and Bastani, Osbert and Sanketi, Pannag R. and Miller, Patrick "Tree" and Yin, Patrick and Wohlhart, Paul and Xu, Peng and Fagan, Peter David and Mitrano, Peter and Sermanet, Pierre and Abbeel, Pieter and Sundaresan, Priya and Chen, Qiuyu and Vuong, Quan and Rafailov, Rafael and Tian, Ran and Doshi, Ria and {Mart{\'i}n-Mart{\'i}n}, Roberto and Baijal, Rohan and Scalise, Rosario and Hendrix, Rose and Lin, Roy and Qian, Runjia and Zhang, Ruohan and Mendonca, Russell and Shah, Rutav and Hoque, Ryan and Julian, Ryan and Bustamante, Samuel and Kirmani, Sean and Levine, Sergey and Lin, Shan and Moore, Sherry and Bahl, Shikhar and Dass, Shivin and Sonawani, Shubham and Tulsiani, Shubham and Song, Shuran and Xu, Sichun and Haldar, Siddhant and Karamcheti, Siddharth and Adebola, Simeon and Guist, Simon and Nasiriany, Soroush and Schaal, Stefan and Welker, Stefan and Tian, Stephen and Ramamoorthy, Subramanian and Dasari, Sudeep and Belkhale, Suneel and Park, Sungjae and Nair, Suraj and Mirchandani, Suvir and Osa, Takayuki and Gupta, Tanmay and Harada, Tatsuya and Matsushima, Tatsuya and Xiao, Ted and Kollar, Thomas and Yu, Tianhe and Ding, Tianli and Davchev, Todor and Zhao, Tony Z. and Armstrong, Travis and Darrell, Trevor and Chung, Trinity and Jain, Vidhi and Kumar, Vikash and Vanhoucke, Vincent and Guizilini, Vitor and Zhan, Wei and Zhou, Wenxuan and Burgard, Wolfram and Chen, Xi and Chen, Xiangyu and Wang, Xiaolong and Zhu, Xinghao and Geng, Xinyang and Liu, Xiyuan and Liangwei, Xu and Li, Xuanlin and Pang, Yansong and Lu, Yao and Ma, Yecheng Jason and Kim, Yejin and Chebotar, Yevgen and Zhou, Yifan and Zhu, Yifeng and Wu, Yilin and Xu, Ying and Wang, Yixuan and Bisk, Yonatan and Dou, Yongqiang and Cho, Yoonyoung and Lee, Youngwoon and Cui, Yuchen and Cao, Yue and Wu, Yueh-Hua and Tang, Yujin and Zhu, Yuke and Zhang, Yunchu and Jiang, Yunfan and Li, Yunshuang and Li, Yunzhu and Iwasawa, Yusuke and Matsuo, Yutaka and Ma, Zehan and Xu, Zhuo and Cui, Zichen Jeff and Zhang, Zichen and Fu, Zipeng and Lin, Zipeng},448 year = {2025},449 month = may,450 number = {arXiv:2310.08864},451 eprint = {2310.08864},452 primaryclass = {cs},453 publisher = {arXiv},454 doi = {10.48550/arXiv.2310.08864},455 urldate = {2025-09-08},456 abstract = {Large, high-capacity models trained on diverse datasets have shown remarkable successes on efficiently tackling downstream applications. In domains from NLP to Computer Vision, this has led to a consolidation of pretrained models, with general pretrained backbones serving as a starting point for many applications. Can such a consolidation happen in robotics? Conventionally, robotic learning methods train a separate model for every application, every robot, and even every environment. Can we instead train generalist X-robot policy that can be adapted efficiently to new robots, tasks, and environments? In this paper, we provide datasets in standardized data formats and models to make it possible to explore this possibility in the context of robotic manipulation, alongside experimental results that provide an example of effective X-robot policies. We assemble a dataset from 22 different robots collected through a collaboration between 21 institutions, demonstrating 527 skills (160266 tasks). We show that a high-capacity model trained on this data, which we call RT-X, exhibits positive transfer and improves the capabilities of multiple robots by leveraging experience from other platforms. More details can be found on the project website https://robotics-transformer-x.github.io.},457 archiveprefix = {arXiv},458 keywords = {Computer Science - Robotics},459 file = {/Users/fracapuano/Zotero/storage/2U73MMVN/Collaboration et al. - 2025 - Open X-Embodiment Robotic Learning Datasets and RT-X Models.pdf;/Users/fracapuano/Zotero/storage/PX7IHY32/2310.html}460}461 462@book{connellRobotLearning1993,463 title = {Robot {Learning}},464 editor = {Connell, Jonathan H. and Mahadevan, Sridhar},465 year = {1993},466 publisher = {Springer US},467 address = {Boston, MA},468 doi = {10.1007/978-1-4615-3184-5},469 urldate = {2025-08-28},470 copyright = {http://www.springer.com/tdm},471 isbn = {978-1-4613-6396-5 978-1-4615-3184-5},472 keywords = {algorithms,artificial intelligence,artificial life,autonom,autonomous robot,genetic algorithms,intelligence,learning,Navigation,programming,proving,robot,uncertainty}473}474 475@article{degraveMagneticControlTokamak2022,476 title = {Magnetic Control of Tokamak Plasmas through Deep Reinforcement Learning},477 author = {Degrave, Jonas and Felici, Federico and Buchli, Jonas and Neunert, Michael and Tracey, Brendan and Carpanese, Francesco and Ewalds, Timo and Hafner, Roland and Abdolmaleki, Abbas and {de las Casas}, Diego and Donner, Craig and Fritz, Leslie and Galperti, Cristian and Huber, Andrea and Keeling, James and Tsimpoukelli, Maria and Kay, Jackie and Merle, Antoine and Moret, Jean-Marc and Noury, Seb and Pesamosca, Federico and Pfau, David and Sauter, Olivier and Sommariva, Cristian and Coda, Stefano and Duval, Basil and Fasoli, Ambrogio and Kohli, Pushmeet and Kavukcuoglu, Koray and Hassabis, Demis and Riedmiller, Martin},478 year = {2022},479 month = feb,480 journal = {Nature},481 volume = {602},482 number = {7897},483 pages = {414--419},484 publisher = {Nature Publishing Group},485 issn = {1476-4687},486 doi = {10.1038/s41586-021-04301-9},487 urldate = {2025-08-31},488 abstract = {Nuclear fusion using magnetic confinement, in particular in the tokamak configuration, is a promising path towards sustainable energy. A core challenge is to shape and maintain a high-temperature plasma within the tokamak vessel. This requires high-dimensional, high-frequency, closed-loop control using magnetic actuator coils, further complicated by the diverse requirements across a wide range of plasma configurations. In this work, we introduce a previously undescribed architecture for tokamak magnetic controller design that autonomously learns to command the full set of control coils. This architecture meets control objectives specified at a high level, at the same time satisfying physical and operational constraints. This approach has unprecedented flexibility and generality in problem specification and yields a notable reduction in design effort to produce new plasma configurations. We successfully produce and control a diverse set of plasma configurations on the Tokamak {\`a} Configuration Variable1,2, including elongated, conventional shapes, as well as advanced configurations, such as negative triangularity and `snowflake' configurations. Our approach achieves accurate tracking of the location, current and shape for these configurations. We also demonstrate sustained `droplets' on TCV, in which two separate plasmas are maintained simultaneously within the vessel. This represents a notable advance for tokamak feedback control, showing the potential of reinforcement learning to accelerate research in the fusion domain, and is one of the most challenging real-world systems to which reinforcement learning has been applied.},489 copyright = {2022 The Author(s)},490 langid = {english},491 keywords = {Computer science,Magnetically confined plasmas,Nuclear fusion and fission},492 file = {/Users/fracapuano/Zotero/storage/EZ4EAU84/Degrave et al. - 2022 - Magnetic control of tokamak plasmas through deep reinforcement learning.pdf}493}494 495@misc{devlinBERTPretrainingDeep2019,496 title = {BERT: Pre-training of {Deep Bidirectional Transformers} for {Language Understanding}},497 shorttitle = {BERT},498 author = {Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina},499 year = {2019},500 month = may,501 number = {arXiv:1810.04805},502 eprint = {1810.04805},503 primaryclass = {cs},504 publisher = {arXiv},505 doi = {10.48550/arXiv.1810.04805},506 urldate = {2025-09-08},507 abstract = {We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement).},508 archiveprefix = {arXiv},509 keywords = {Computer Science - Computation and Language},510 file = {/Users/fracapuano/Zotero/storage/AJ3SRLHF/Devlin et al. - 2019 - BERT Pre-training of Deep Bidirectional Transformers for Language Understanding.pdf;/Users/fracapuano/Zotero/storage/LNIKJNIW/1810.html}511}512 513@misc{driessKnowledgeInsulatingVisionLanguageAction2025,514 title = {Knowledge Insulating Vision-Language-Action Models: {Train Fast}, {Run Fast}, {Generalize Better}},515 shorttitle = {Knowledge Insulating Vision-Language-Action Models},516 author = {Driess, Danny and Springenberg, Jost Tobias and Ichter, Brian and Yu, Lili and {Li-Bell}, Adrian and Pertsch, Karl and Ren, Allen Z. and Walke, Homer and Vuong, Quan and Shi, Lucy Xiaoyang and Levine, Sergey},517 year = {2025},518 month = may,519 number = {arXiv:2505.23705},520 eprint = {2505.23705},521 primaryclass = {cs},522 publisher = {arXiv},523 doi = {10.48550/arXiv.2505.23705},524 urldate = {2025-09-09},525 abstract = {Vision-language-action (VLA) models provide a powerful approach to training control policies for physical systems, such as robots, by combining end-to-end learning with transfer of semantic knowledge from web-scale vision-language model (VLM) training. However, the constraints of real-time control are often at odds with the design of VLMs: the most powerful VLMs have tens or hundreds of billions of parameters, presenting an obstacle to real-time inference, and operate on discrete tokens rather than the continuous-valued outputs that are required for controlling robots. To address this challenge, recent VLA models have used specialized modules for efficient continuous control, such as action experts or continuous output heads, which typically require adding new untrained parameters to the pretrained VLM backbone. While these modules improve real-time and control capabilities, it remains an open question whether they preserve or degrade the semantic knowledge contained in the pretrained VLM, and what effect they have on the VLA training dynamics. In this paper, we study this question in the context of VLAs that include a continuous diffusion or flow matching action expert, showing that naively including such experts significantly harms both training speed and knowledge transfer. We provide an extensive analysis of various design choices, their impact on performance and knowledge transfer, and propose a technique for insulating the VLM backbone during VLA training that mitigates this issue. Videos are available at https://pi.website/research/knowledge_insulation.},526 archiveprefix = {arXiv},527 keywords = {Computer Science - Machine Learning,Computer Science - Robotics},528 file = {/Users/fracapuano/Zotero/storage/QHTS9JIC/Driess et al. - 2025 - Knowledge Insulating Vision-Language-Action Models Train Fast, Run Fast, Generalize Better.pdf;/Users/fracapuano/Zotero/storage/3U9FCXRB/2505.html}529}530 531@misc{driessPaLMEEmbodiedMultimodal2023,532 title = {PaLM-E: {An Embodied Multimodal Language Model}},533 shorttitle = {PaLM-E},534 author = {Driess, Danny and Xia, Fei and Sajjadi, Mehdi S. M. and Lynch, Corey and Chowdhery, Aakanksha and Ichter, Brian and Wahid, Ayzaan and Tompson, Jonathan and Vuong, Quan and Yu, Tianhe and Huang, Wenlong and Chebotar, Yevgen and Sermanet, Pierre and Duckworth, Daniel and Levine, Sergey and Vanhoucke, Vincent and Hausman, Karol and Toussaint, Marc and Greff, Klaus and Zeng, Andy and Mordatch, Igor and Florence, Pete},535 year = {2023},536 month = mar,537 number = {arXiv:2303.03378},538 eprint = {2303.03378},539 primaryclass = {cs},540 publisher = {arXiv},541 doi = {10.48550/arXiv.2303.03378},542 urldate = {2025-09-07},543 abstract = {Large language models excel at a wide range of complex tasks. However, enabling general inference in the real world, e.g., for robotics problems, raises the challenge of grounding. We propose embodied language models to directly incorporate real-world continuous sensor modalities into language models and thereby establish the link between words and percepts. Input to our embodied language model are multi-modal sentences that interleave visual, continuous state estimation, and textual input encodings. We train these encodings end-to-end, in conjunction with a pre-trained large language model, for multiple embodied tasks including sequential robotic manipulation planning, visual question answering, and captioning. Our evaluations show that PaLM-E, a single large embodied multimodal model, can address a variety of embodied reasoning tasks, from a variety of observation modalities, on multiple embodiments, and further, exhibits positive transfer: the model benefits from diverse joint training across internet-scale language, vision, and visual-language domains. Our largest model, PaLM-E-562B with 562B parameters, in addition to being trained on robotics tasks, is a visual-language generalist with state-of-the-art performance on OK-VQA, and retains generalist language capabilities with increasing scale.},544 archiveprefix = {arXiv},545 keywords = {Computer Science - Artificial Intelligence,Computer Science - Machine Learning,Computer Science - Robotics},546 file = {/Users/fracapuano/Zotero/storage/PQSPI784/Driess et al. - 2023 - PaLM-E An Embodied Multimodal Language Model.pdf;/Users/fracapuano/Zotero/storage/K3PJVSGB/2303.html}547}548 549@misc{esserScalingRectifiedFlow2024,550 title = {Scaling {Rectified Flow Transformers} for High-Resolution Image Synthesis},551 author = {Esser, Patrick and Kulal, Sumith and Blattmann, Andreas and Entezari, Rahim and M{\"u}ller, Jonas and Saini, Harry and Levi, Yam and Lorenz, Dominik and Sauer, Axel and Boesel, Frederic and Podell, Dustin and Dockhorn, Tim and English, Zion and Lacey, Kyle and Goodwin, Alex and Marek, Yannik and Rombach, Robin},552 year = {2024},553 month = mar,554 number = {arXiv:2403.03206},555 eprint = {2403.03206},556 primaryclass = {cs},557 publisher = {arXiv},558 doi = {10.48550/arXiv.2403.03206},559 urldate = {2025-09-07},560 abstract = {Diffusion models create data from noise by inverting the forward paths of data towards noise and have emerged as a powerful generative modeling technique for high-dimensional, perceptual data such as images and videos. Rectified flow is a recent generative model formulation that connects data and noise in a straight line. Despite its better theoretical properties and conceptual simplicity, it is not yet decisively established as standard practice. In this work, we improve existing noise sampling techniques for training rectified flow models by biasing them towards perceptually relevant scales. Through a large-scale study, we demonstrate the superior performance of this approach compared to established diffusion formulations for high-resolution text-to-image synthesis. Additionally, we present a novel transformer-based architecture for text-to-image generation that uses separate weights for the two modalities and enables a bidirectional flow of information between image and text tokens, improving text comprehension, typography, and human preference ratings. We demonstrate that this architecture follows predictable scaling trends and correlates lower validation loss to improved text-to-image synthesis as measured by various metrics and human evaluations. Our largest models outperform state-of-the-art models, and we will make our experimental data, code, and model weights publicly available.},561 archiveprefix = {arXiv},562 keywords = {Computer Science - Computer Vision and Pattern Recognition},563 file = {/Users/fracapuano/Zotero/storage/23TGK9JM/Esser et al. - 2024 - Scaling Rectified Flow Transformers for High-Resolution Image Synthesis.pdf;/Users/fracapuano/Zotero/storage/W2CRYPZY/2403.html}564}565 566@misc{fedusReviewSparseExpert2022,567 title = {A {Review} of {Sparse Expert Models} in {Deep Learning}},568 author = {Fedus, William and Dean, Jeff and Zoph, Barret},569 year = {2022},570 month = sep,571 number = {arXiv:2209.01667},572 eprint = {2209.01667},573 primaryclass = {cs},574 publisher = {arXiv},575 doi = {10.48550/arXiv.2209.01667},576 urldate = {2025-09-08},577 abstract = {Sparse expert models are a thirty-year old concept re-emerging as a popular architecture in deep learning. This class of architecture encompasses Mixture-of-Experts, Switch Transformers, Routing Networks, BASE layers, and others, all with the unifying idea that each example is acted on by a subset of the parameters. By doing so, the degree of sparsity decouples the parameter count from the compute per example allowing for extremely large, but efficient models. The resulting models have demonstrated significant improvements across diverse domains such as natural language processing, computer vision, and speech recognition. We review the concept of sparse expert models, provide a basic description of the common algorithms, contextualize the advances in the deep learning era, and conclude by highlighting areas for future work.},578 archiveprefix = {arXiv},579 keywords = {Computer Science - Computation and Language,Computer Science - Machine Learning},580 file = {/Users/fracapuano/Zotero/storage/MZXG2WMJ/Fedus et al. - 2022 - A Review of Sparse Expert Models in Deep Learning.pdf;/Users/fracapuano/Zotero/storage/GLZINJYC/2209.html}581}582 583@misc{finiMultimodalAutoregressivePretraining2024,584 title = {Multimodal Autoregressive Pre-training of {Large Vision Encoders}},585 author = {Fini, Enrico and Shukor, Mustafa and Li, Xiujun and Dufter, Philipp and Klein, Michal and Haldimann, David and Aitharaju, Sai and da Costa, Victor Guilherme Turrisi and B{\'e}thune, Louis and Gan, Zhe and Toshev, Alexander T. and Eichner, Marcin and Nabi, Moin and Yang, Yinfei and Susskind, Joshua M. and {El-Nouby}, Alaaeldin},586 year = {2024},587 month = nov,588 number = {arXiv:2411.14402},589 eprint = {2411.14402},590 primaryclass = {cs},591 publisher = {arXiv},592 doi = {10.48550/arXiv.2411.14402},593 urldate = {2025-09-09},594 abstract = {We introduce a novel method for pre-training of large-scale vision encoders. Building on recent advancements in autoregressive pre-training of vision models, we extend this framework to a multimodal setting, i.e., images and text. In this paper, we present AIMV2, a family of generalist vision encoders characterized by a straightforward pre-training process, scalability, and remarkable performance across a range of downstream tasks. This is achieved by pairing the vision encoder with a multimodal decoder that autoregressively generates raw image patches and text tokens. Our encoders excel not only in multimodal evaluations but also in vision benchmarks such as localization, grounding, and classification. Notably, our AIMV2-3B encoder achieves 89.5% accuracy on ImageNet-1k with a frozen trunk. Furthermore, AIMV2 consistently outperforms state-of-the-art contrastive models (e.g., CLIP, SigLIP) in multimodal image understanding across diverse settings.},595 archiveprefix = {arXiv},596 keywords = {Computer Science - Computer Vision and Pattern Recognition,Computer Science - Machine Learning},597 file = {/Users/fracapuano/Zotero/storage/ULTX55I6/Fini et al. - 2024 - Multimodal Autoregressive Pre-training of Large Vision Encoders.pdf;/Users/fracapuano/Zotero/storage/SUG2W6A9/2411.html}598}599 600@inproceedings{florenceImplicitBehavioralCloning2022,601 title = {Implicit {Behavioral Cloning}},602 booktitle = {Proceedings of the 5th {Conference} on {Robot Learning}},603 author = {Florence, Pete and Lynch, Corey and Zeng, Andy and Ramirez, Oscar A. and Wahid, Ayzaan and Downs, Laura and Wong, Adrian and Lee, Johnny and Mordatch, Igor and Tompson, Jonathan},604 year = {2022},605 month = jan,606 pages = {158--168},607 publisher = {PMLR},608 issn = {2640-3498},609 urldate = {2025-09-01},610 abstract = {We find that across a wide range of robot policy learning scenarios, treating supervised policy learning with an implicit model generally performs better, on average, than commonly used explicit models. We present extensive experiments on this finding, and we provide both intuitive insight and theoretical arguments distinguishing the properties of implicit models compared to their explicit counterparts, particularly with respect to approximating complex, potentially discontinuous and multi-valued (set-valued) functions. On robotic policy learning tasks we show that implicit behavior-cloning policies with energy-based models (EBM) often outperform common explicit (Mean Square Error, or Mixture Density) behavior-cloning policies, including on tasks with high-dimensional action spaces and visual image inputs. We find these policies provide competitive results or outperform state-of-the-art offline reinforcement learning methods on the challenging human-expert tasks from the D4RL benchmark suite, despite using no reward information. In the real world, robots with implicit policies can learn complex and remarkably subtle behaviors on contact-rich tasks from human demonstrations, including tasks with high combinatorial complexity and tasks requiring 1mm precision.},611 langid = {english},612 file = {/Users/fracapuano/Zotero/storage/Q8I5E862/Florence et al. - 2022 - Implicit Behavioral Cloning.pdf}613}614 615@misc{FROMAGe,616 title = {Grounding Language Models to Images for Multimodal Inputs and Outputs},617 author = {Koh, Jing Yu and Salakhutdinov, Ruslan and Fried, Daniel},618 year = {2023}619}620 621@article{fujitaDevelopmentRobotsNuclear2020,622 title = {Development of {Robots} for {Nuclear Power Plants} and {Their Application} to {New Fields}},623 author = {Fujita, Jun and Soda, Daisuke and Murata, Chotaro and Tsuhari, Hiroyuki},624 year = {2020},625 volume = {57},626 number = {4},627 langid = {english},628 file = {/Users/fracapuano/Zotero/storage/K349QTEG/Fujita et al. - 2020 - Development of Robots for Nuclear Power Plants and Their Application to New Fields.pdf}629}630 631@misc{grattafioriLlama3Herd2024,632 title = {The {Llama} 3 {Herd} of {Models}},633 author = {Grattafiori, Aaron and Dubey, Abhimanyu and Jauhri, Abhinav and Pandey, Abhinav and Kadian, Abhishek and {Al-Dahle}, Ahmad and Letman, Aiesha and Mathur, Akhil and Schelten, Alan and Vaughan, Alex and Yang, Amy and Fan, Angela and Goyal, Anirudh and Hartshorn, Anthony and Yang, Aobo and Mitra, Archi and Sravankumar, Archie and Korenev, Artem and Hinsvark, Arthur and Rao, Arun and Zhang, Aston and Rodriguez, Aurelien and Gregerson, Austen and Spataru, Ava and Roziere, Baptiste and Biron, Bethany and Tang, Binh and Chern, Bobbie and Caucheteux, Charlotte and Nayak, Chaya and Bi, Chloe and Marra, Chris and McConnell, Chris and Keller, Christian and Touret, Christophe and Wu, Chunyang and Wong, Corinne and Ferrer, Cristian Canton and Nikolaidis, Cyrus and Allonsius, Damien and Song, Daniel and Pintz, Danielle and Livshits, Danny and Wyatt, Danny and Esiobu, David and Choudhary, Dhruv and Mahajan, Dhruv and {Garcia-Olano}, Diego and Perino, Diego and Hupkes, Dieuwke and Lakomkin, Egor and AlBadawy, Ehab and Lobanova, Elina and Dinan, Emily and Smith, Eric Michael and Radenovic, Filip and Guzm{\'a}n, Francisco and Zhang, Frank and Synnaeve, Gabriel and Lee, Gabrielle and Anderson, Georgia Lewis and Thattai, Govind and Nail, Graeme and Mialon, Gregoire and Pang, Guan and Cucurell, Guillem and Nguyen, Hailey and Korevaar, Hannah and Xu, Hu and Touvron, Hugo and Zarov, Iliyan and Ibarra, Imanol Arrieta and Kloumann, Isabel and Misra, Ishan and Evtimov, Ivan and Zhang, Jack and Copet, Jade and Lee, Jaewon and Geffert, Jan and Vranes, Jana and Park, Jason and Mahadeokar, Jay and Shah, Jeet and van der Linde, Jelmer and Billock, Jennifer and Hong, Jenny and Lee, Jenya and Fu, Jeremy and Chi, Jianfeng and Huang, Jianyu and Liu, Jiawen and Wang, Jie and Yu, Jiecao and Bitton, Joanna and Spisak, Joe and Park, Jongsoo and Rocca, Joseph and Johnstun, Joshua and Saxe, Joshua and Jia, Junteng and Alwala, Kalyan Vasuden and Prasad, Karthik and Upasani, Kartikeya and Plawiak, Kate and Li, Ke and Heafield, Kenneth and Stone, Kevin and {El-Arini}, Khalid and Iyer, Krithika and Malik, Kshitiz and Chiu, Kuenley and Bhalla, Kunal and Lakhotia, Kushal and {Rantala-Yeary}, Lauren and van der Maaten, Laurens and Chen, Lawrence and Tan, Liang and Jenkins, Liz and Martin, Louis and Madaan, Lovish and Malo, Lubo and Blecher, Lukas and Landzaat, Lukas and de Oliveira, Luke and Muzzi, Madeline and Pasupuleti, Mahesh and Singh, Mannat and Paluri, Manohar and Kardas, Marcin and Tsimpoukelli, Maria and Oldham, Mathew and Rita, Mathieu and Pavlova, Maya and Kambadur, Melanie and Lewis, Mike and Si, Min and Singh, Mitesh Kumar and Hassan, Mona and Goyal, Naman and Torabi, Narjes and Bashlykov, Nikolay and Bogoychev, Nikolay and Chatterji, Niladri and Zhang, Ning and Duchenne, Olivier and {\c C}elebi, Onur and Alrassy, Patrick and Zhang, Pengchuan and Li, Pengwei and Vasic, Petar and Weng, Peter and Bhargava, Prajjwal and Dubal, Pratik and Krishnan, Praveen and Koura, Punit Singh and Xu, Puxin and He, Qing and Dong, Qingxiao and Srinivasan, Ragavan and Ganapathy, Raj and Calderer, Ramon and Cabral, Ricardo Silveira and Stojnic, Robert and Raileanu, Roberta and Maheswari, Rohan and Girdhar, Rohit and Patel, Rohit and Sauvestre, Romain and Polidoro, Ronnie and Sumbaly, Roshan and Taylor, Ross and Silva, Ruan and Hou, Rui and Wang, Rui and Hosseini, Saghar and Chennabasappa, Sahana and Singh, Sanjay and Bell, Sean and Kim, Seohyun Sonia and Edunov, Sergey and Nie, Shaoliang and Narang, Sharan and Raparthy, Sharath and Shen, Sheng and Wan, Shengye and Bhosale, Shruti and Zhang, Shun and Vandenhende, Simon and Batra, Soumya and Whitman, Spencer and Sootla, Sten and Collot, Stephane and Gururangan, Suchin and Borodinsky, Sydney and Herman, Tamar and Fowler, Tara and Sheasha, Tarek and Georgiou, Thomas and Scialom, Thomas and Speckbacher, Tobias and Mihaylov, Todor and Xiao, Tong and Karn, Ujjwal and Goswami, Vedanuj and Gupta, Vibhor and Ramanathan, Vignesh and Kerkez, Viktor and Gonguet, Vincent and Do, Virginie and Vogeti, Vish and Albiero, V{\'i}tor and Petrovic, Vladan and Chu, Weiwei and Xiong, Wenhan and Fu, Wenyin and Meers, Whitney and Martinet, Xavier and Wang, Xiaodong and Wang, Xiaofang and Tan, Xiaoqing Ellen and Xia, Xide and Xie, Xinfeng and Jia, Xuchao and Wang, Xuewei and Goldschlag, Yaelle and Gaur, Yashesh and Babaei, Yasmine and Wen, Yi and Song, Yiwen and Zhang, Yuchen and Li, Yue and Mao, Yuning and Coudert, Zacharie Delpierre and Yan, Zheng and Chen, Zhengxing and Papakipos, Zoe and Singh, Aaditya and Srivastava, Aayushi and Jain, Abha and Kelsey, Adam and Shajnfeld, Adam and Gangidi, Adithya and Victoria, Adolfo and Goldstand, Ahuva and Menon, Ajay and Sharma, Ajay and Boesenberg, Alex and Baevski, Alexei and Feinstein, Allie and Kallet, Amanda and Sangani, Amit and Teo, Amos and Yunus, Anam and Lupu, Andrei and Alvarado, Andres and Caples, Andrew and Gu, Andrew and Ho, Andrew and Poulton, Andrew and Ryan, Andrew and Ramchandani, Ankit and Dong, Annie and Franco, Annie and Goyal, Anuj and Saraf, Aparajita and Chowdhury, Arkabandhu and Gabriel, Ashley and Bharambe, Ashwin and Eisenman, Assaf and Yazdan, Azadeh and James, Beau and Maurer, Ben and Leonhardi, Benjamin and Huang, Bernie and Loyd, Beth and Paola, Beto De and Paranjape, Bhargavi and Liu, Bing and Wu, Bo and Ni, Boyu and Hancock, Braden and Wasti, Bram and Spence, Brandon and Stojkovic, Brani and Gamido, Brian and Montalvo, Britt and Parker, Carl and Burton, Carly and Mejia, Catalina and Liu, Ce and Wang, Changhan and Kim, Changkyu and Zhou, Chao and Hu, Chester and Chu, Ching-Hsiang and Cai, Chris and Tindal, Chris and Feichtenhofer, Christoph and Gao, Cynthia and Civin, Damon and Beaty, Dana and Kreymer, Daniel and Li, Daniel and Adkins, David and Xu, David and Testuggine, Davide and David, Delia and Parikh, Devi and Liskovich, Diana and Foss, Didem and Wang, Dingkang and Le, Duc and Holland, Dustin and Dowling, Edward and Jamil, Eissa and Montgomery, Elaine and Presani, Eleonora and Hahn, Emily and Wood, Emily and Le, Eric-Tuan and Brinkman, Erik and Arcaute, Esteban and Dunbar, Evan and Smothers, Evan and Sun, Fei and Kreuk, Felix and Tian, Feng and Kokkinos, Filippos and Ozgenel, Firat and Caggioni, Francesco and Kanayet, Frank and Seide, Frank and Florez, Gabriela Medina and Schwarz, Gabriella and Badeer, Gada and Swee, Georgia and Halpern, Gil and Herman, Grant and Sizov, Grigory and Guangyi and Zhang and Lakshminarayanan, Guna and Inan, Hakan and Shojanazeri, Hamid and Zou, Han and Wang, Hannah and Zha, Hanwen and Habeeb, Haroun and Rudolph, Harrison and Suk, Helen and Aspegren, Henry and Goldman, Hunter and Zhan, Hongyuan and Damlaj, Ibrahim and Molybog, Igor and Tufanov, Igor and Leontiadis, Ilias and Veliche, Irina-Elena and Gat, Itai and Weissman, Jake and Geboski, James and Kohli, James and Lam, Janice and Asher, Japhet and Gaya, Jean-Baptiste and Marcus, Jeff and Tang, Jeff and Chan, Jennifer and Zhen, Jenny and Reizenstein, Jeremy and Teboul, Jeremy and Zhong, Jessica and Jin, Jian and Yang, Jingyi and Cummings, Joe and Carvill, Jon and Shepard, Jon and McPhie, Jonathan and Torres, Jonathan and Ginsburg, Josh and Wang, Junjie and Wu, Kai and U, Kam Hou and Saxena, Karan and Khandelwal, Kartikay and Zand, Katayoun and Matosich, Kathy and Veeraraghavan, Kaushik and Michelena, Kelly and Li, Keqian and Jagadeesh, Kiran and Huang, Kun and Chawla, Kunal and Huang, Kyle and Chen, Lailin and Garg, Lakshya and A, Lavender and Silva, Leandro and Bell, Lee and Zhang, Lei and Guo, Liangpeng and Yu, Licheng and Moshkovich, Liron and Wehrstedt, Luca and Khabsa, Madian and Avalani, Manav and Bhatt, Manish and Mankus, Martynas and Hasson, Matan and Lennie, Matthew and Reso, Matthias and Groshev, Maxim and Naumov, Maxim and Lathi, Maya and Keneally, Meghan and Liu, Miao and Seltzer, Michael L. and Valko, Michal and Restrepo, Michelle and Patel, Mihir and Vyatskov, Mik and Samvelyan, Mikayel and Clark, Mike and Macey, Mike and Wang, Mike and Hermoso, Miquel Jubert and Metanat, Mo and Rastegari, Mohammad and Bansal, Munish and Santhanam, Nandhini and Parks, Natascha and White, Natasha and Bawa, Navyata and Singhal, Nayan and Egebo, Nick and Usunier, Nicolas and Mehta, Nikhil and Laptev, Nikolay Pavlovich and Dong, Ning and Cheng, Norman and Chernoguz, Oleg and Hart, Olivia and Salpekar, Omkar and Kalinli, Ozlem and Kent, Parkin and Parekh, Parth and Saab, Paul and Balaji, Pavan and Rittner, Pedro and Bontrager, Philip and Roux, Pierre and Dollar, Piotr and Zvyagina, Polina and Ratanchandani, Prashant and Yuvraj, Pritish and Liang, Qian and Alao, Rachad and Rodriguez, Rachel and Ayub, Rafi and Murthy, Raghotham and Nayani, Raghu and Mitra, Rahul and Parthasarathy, Rangaprabhu and Li, Raymond and Hogan, Rebekkah and Battey, Robin and Wang, Rocky and Howes, Russ and Rinott, Ruty and Mehta, Sachin and Siby, Sachin and Bondu, Sai Jayesh and Datta, Samyak and Chugh, Sara and Hunt, Sara and Dhillon, Sargun and Sidorov, Sasha and Pan, Satadru and Mahajan, Saurabh and Verma, Saurabh and Yamamoto, Seiji and Ramaswamy, Sharadh and Lindsay, Shaun and Lindsay, Shaun and Feng, Sheng and Lin, Shenghao and Zha, Shengxin Cindy and Patil, Shishir and Shankar, Shiva and Zhang, Shuqiang and Zhang, Shuqiang and Wang, Sinong and Agarwal, Sneha and Sajuyigbe, Soji and Chintala, Soumith and Max, Stephanie and Chen, Stephen and Kehoe, Steve and Satterfield, Steve and Govindaprasad, Sudarshan and Gupta, Sumit and Deng, Summer and Cho, Sungmin and Virk, Sunny and Subramanian, Suraj and Choudhury, Sy and Goldman, Sydney and Remez, Tal and Glaser, Tamar and Best, Tamara and Koehler, Thilo and Robinson, Thomas and Li, Tianhe and Zhang, Tianjun and Matthews, Tim and Chou, Timothy and Shaked, Tzook and Vontimitta, Varun and Ajayi, Victoria and Montanez, Victoria and Mohan, Vijai and Kumar, Vinay Satish and Mangla, Vishal and Ionescu, Vlad and Poenaru, Vlad and Mihailescu, Vlad Tiberiu and Ivanov, Vladimir and Li, Wei and Wang, Wenchen and Jiang, Wenwen and Bouaziz, Wes and Constable, Will and Tang, Xiaocheng and Wu, Xiaojian and Wang, Xiaolan and Wu, Xilun and Gao, Xinbo and Kleinman, Yaniv and Chen, Yanjun and Hu, Ye and Jia, Ye and Qi, Ye and Li, Yenda and Zhang, Yilin and Zhang, Ying and Adi, Yossi and Nam, Youngjin and Yu and Wang and Zhao, Yu and Hao, Yuchen and Qian, Yundi and Li, Yunlu and He, Yuzi and Rait, Zach and DeVito, Zachary and Rosnbrick, Zef and Wen, Zhaoduo and Yang, Zhenyu and Zhao, Zhiwei and Ma, Zhiyu},634 year = {2024},635 month = nov,636 number = {arXiv:2407.21783},637 eprint = {2407.21783},638 primaryclass = {cs},639 publisher = {arXiv},640 doi = {10.48550/arXiv.2407.21783},641 urldate = {2025-09-09},642 abstract = {Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.},643 archiveprefix = {arXiv},644 keywords = {Computer Science - Artificial Intelligence,Computer Science - Computation and Language,Computer Science - Computer Vision and Pattern Recognition},645 file = {/Users/fracapuano/Zotero/storage/88PJ48EN/Grattafiori et al. - 2024 - The Llama 3 Herd of Models.pdf;/Users/fracapuano/Zotero/storage/2LLAWX8L/2407.html}646}647 648@inproceedings{griffinWalkingStabilizationUsing2017,649 title = {Walking {Stabilization Using Step Timing} and {Location Adjustment} on the {Humanoid Robot}, {Atlas}},650 booktitle = {2017 {IEEE}/RSJ International Conference on {Intelligent Robots} and {Systems} ({IROS})},651 author = {Griffin, Robert J. and Wiedebach, Georg and Bertrand, Sylvain and Leonessa, Alexander and Pratt, Jerry},652 year = {2017},653 month = sep,654 eprint = {1703.00477},655 primaryclass = {cs},656 pages = {667--673},657 doi = {10.1109/IROS.2017.8202223},658 urldate = {2025-08-26},659 abstract = {While humans are highly capable of recovering from external disturbances and uncertainties that result in large tracking errors, humanoid robots have yet to reliably mimic this level of robustness. Essential to this is the ability to combine traditional "ankle strategy" balancing with step timing and location adjustment techniques. In doing so, the robot is able to step quickly to the necessary location to continue walking. In this work, we present both a new swing speed up algorithm to adjust the step timing, allowing the robot to set the foot down more quickly to recover from errors in the direction of the current capture point dynamics, and a new algorithm to adjust the desired footstep, expanding the base of support to utilize the center of pressure (CoP)-based ankle strategy for balance. We then utilize the desired centroidal moment pivot (CMP) to calculate the momentum rate of change for our inverse-dynamics based whole-body controller. We present simulation and experimental results using this work, and discuss performance limitations and potential improvements.},660 archiveprefix = {arXiv},661 keywords = {Computer Science - Robotics},662 file = {/Users/fracapuano/Zotero/storage/SSNAZ6U4/Griffin et al. - 2017 - Walking Stabilization Using Step Timing and Location Adjustment on the Humanoid Robot, Atlas.pdf;/Users/fracapuano/Zotero/storage/VP885PA9/1703.html}663}664 665@misc{haarnojaReinforcementLearningDeep2017,666 title = {Reinforcement {Learning} with Deep Energy-Based Policies},667 author = {Haarnoja, Tuomas and Tang, Haoran and Abbeel, Pieter and Levine, Sergey},668 year = {2017},669 month = jul,670 number = {arXiv:1702.08165},671 eprint = {1702.08165},672 primaryclass = {cs},673 publisher = {arXiv},674 doi = {10.48550/arXiv.1702.08165},675 urldate = {2025-08-31},676 abstract = {We propose a method for learning expressive energy-based policies for continuous states and actions, which has been feasible only in tabular domains before. We apply our method to learning maximum entropy policies, resulting into a new algorithm, called soft Q-learning, that expresses the optimal policy via a Boltzmann distribution. We use the recently proposed amortized Stein variational gradient descent to learn a stochastic sampling network that approximates samples from this distribution. The benefits of the proposed algorithm include improved exploration and compositionality that allows transferring skills between tasks, which we confirm in simulated experiments with swimming and walking robots. We also draw a connection to actor-critic methods, which can be viewed performing approximate inference on the corresponding energy-based model.},677 archiveprefix = {arXiv},678 keywords = {Computer Science - Artificial Intelligence,Computer Science - Machine Learning},679 file = {/Users/fracapuano/Zotero/storage/PXCR4TCT/Haarnoja et al. - 2017 - Reinforcement Learning with Deep Energy-Based Policies.pdf;/Users/fracapuano/Zotero/storage/VUXXX9B7/1702.html}680}681 682@misc{haarnojaReinforcementLearningDeep2017a,683 title = {Reinforcement {Learning} with Deep Energy-Based Policies},684 author = {Haarnoja, Tuomas and Tang, Haoran and Abbeel, Pieter and Levine, Sergey},685 year = {2017},686 month = jul,687 number = {arXiv:1702.08165},688 eprint = {1702.08165},689 primaryclass = {cs},690 publisher = {arXiv},691 doi = {10.48550/arXiv.1702.08165},692 urldate = {2025-08-31},693 abstract = {We propose a method for learning expressive energy-based policies for continuous states and actions, which has been feasible only in tabular domains before. We apply our method to learning maximum entropy policies, resulting into a new algorithm, called soft Q-learning, that expresses the optimal policy via a Boltzmann distribution. We use the recently proposed amortized Stein variational gradient descent to learn a stochastic sampling network that approximates samples from this distribution. The benefits of the proposed algorithm include improved exploration and compositionality that allows transferring skills between tasks, which we confirm in simulated experiments with swimming and walking robots. We also draw a connection to actor-critic methods, which can be viewed performing approximate inference on the corresponding energy-based model.},694 archiveprefix = {arXiv},695 keywords = {Computer Science - Artificial Intelligence,Computer Science - Machine Learning},696 file = {/Users/fracapuano/Zotero/storage/T84UBYDJ/Haarnoja et al. - 2017 - Reinforcement Learning with Deep Energy-Based Policies.pdf;/Users/fracapuano/Zotero/storage/53SJ2ED8/1702.html}697}698 699@inproceedings{haarnojaReinforcementLearningDeep2017b,700 title = {Reinforcement {Learning} with Deep Energy-Based Policies},701 booktitle = {Proceedings of the 34th {International Conference} on {Machine Learning}},702 author = {Haarnoja, Tuomas and Tang, Haoran and Abbeel, Pieter and Levine, Sergey},703 year = {2017},704 month = jul,705 pages = {1352--1361},706 publisher = {PMLR},707 issn = {2640-3498},708 urldate = {2025-08-31},709 abstract = {We propose a method for learning expressive energy-based policies for continuous states and actions, which has been feasible only in tabular domains before. We apply our method to learning maximum entropy policies, resulting into a new algorithm, called soft Q-learning, that expresses the optimal policy via a Boltzmann distribution. We use the recently proposed amortized Stein variational gradient descent to learn a stochastic sampling network that approximates samples from this distribution. The benefits of the proposed algorithm include improved exploration and compositionality that allows transferring skills between tasks, which we confirm in simulated experiments with swimming and walking robots. We also draw a connection to actor-critic methods, which can be viewed performing approximate inference on the corresponding energy-based model.},710 langid = {english},711 file = {/Users/fracapuano/Zotero/storage/C59BJ4GU/Haarnoja et al. - 2017 - Reinforcement Learning with Deep Energy-Based Policies.pdf}712}713 714@misc{haarnojaSoftActorCriticOffPolicy2018,715 title = {Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a {Stochastic Actor}},716 shorttitle = {Soft Actor-Critic},717 author = {Haarnoja, Tuomas and Zhou, Aurick and Abbeel, Pieter and Levine, Sergey},718 year = {2018},719 month = aug,720 number = {arXiv:1801.01290},721 eprint = {1801.01290},722 primaryclass = {cs},723 publisher = {arXiv},724 doi = {10.48550/arXiv.1801.01290},725 urldate = {2025-08-29},726 abstract = {Model-free deep reinforcement learning (RL) algorithms have been demonstrated on a range of challenging decision making and control tasks. However, these methods typically suffer from two major challenges: very high sample complexity and brittle convergence properties, which necessitate meticulous hyperparameter tuning. Both of these challenges severely limit the applicability of such methods to complex, real-world domains. In this paper, we propose soft actor-critic, an off-policy actor-critic deep RL algorithm based on the maximum entropy reinforcement learning framework. In this framework, the actor aims to maximize expected reward while also maximizing entropy. That is, to succeed at the task while acting as randomly as possible. Prior deep RL methods based on this framework have been formulated as Q-learning methods. By combining off-policy updates with a stable stochastic actor-critic formulation, our method achieves state-of-the-art performance on a range of continuous control benchmark tasks, outperforming prior on-policy and off-policy methods. Furthermore, we demonstrate that, in contrast to other off-policy algorithms, our approach is very stable, achieving very similar performance across different random seeds.},727 archiveprefix = {arXiv},728 keywords = {Computer Science - Artificial Intelligence,Computer Science - Machine Learning,Statistics - Machine Learning},729 file = {/Users/fracapuano/Zotero/storage/HG6UQIRM/Haarnoja et al. - 2018 - Soft Actor-Critic Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor.pdf;/Users/fracapuano/Zotero/storage/RKG3J7MX/1801.html}730}731 732@misc{hansenTemporalDifferenceLearning2022,733 title = {Temporal {Difference Learning} for {Model Predictive Control}},734 author = {Hansen, Nicklas and Wang, Xiaolong and Su, Hao},735 year = {2022},736 month = jul,737 number = {arXiv:2203.04955},738 eprint = {2203.04955},739 primaryclass = {cs},740 publisher = {arXiv},741 doi = {10.48550/arXiv.2203.04955},742 urldate = {2025-08-25},743 abstract = {Data-driven model predictive control has two key advantages over model-free methods: a potential for improved sample efficiency through model learning, and better performance as computational budget for planning increases. However, it is both costly to plan over long horizons and challenging to obtain an accurate model of the environment. In this work, we combine the strengths of model-free and model-based methods. We use a learned task-oriented latent dynamics model for local trajectory optimization over a short horizon, and use a learned terminal value function to estimate long-term return, both of which are learned jointly by temporal difference learning. Our method, TD-MPC, achieves superior sample efficiency and asymptotic performance over prior work on both state and image-based continuous control tasks from DMControl and Meta-World. Code and video results are available at https://nicklashansen.github.io/td-mpc.},744 archiveprefix = {arXiv},745 keywords = {Computer Science - Machine Learning,Computer Science - Robotics},746 file = {/Users/fracapuano/Zotero/storage/TZF8LCDG/Hansen et al. - 2022 - Temporal Difference Learning for Model Predictive Control.pdf;/Users/fracapuano/Zotero/storage/WU2WWWQE/2203.html}747}748 749@misc{heessEmergenceLocomotionBehaviours2017,750 title = {Emergence of {Locomotion Behaviours} in {Rich Environments}},751 author = {Heess, Nicolas and TB, Dhruva and Sriram, Srinivasan and Lemmon, Jay and Merel, Josh and Wayne, Greg and Tassa, Yuval and Erez, Tom and Wang, Ziyu and Eslami, S. M. Ali and Riedmiller, Martin and Silver, David},752 year = {2017},753 month = jul,754 number = {arXiv:1707.02286},755 eprint = {1707.02286},756 primaryclass = {cs},757 publisher = {arXiv},758 doi = {10.48550/arXiv.1707.02286},759 urldate = {2025-09-02},760 abstract = {The reinforcement learning paradigm allows, in principle, for complex behaviours to be learned directly from simple reward signals. In practice, however, it is common to carefully hand-design the reward function to encourage a particular solution, or to derive it from demonstration data. In this paper explore how a rich environment can help to promote the learning of complex behavior. Specifically, we train agents in diverse environmental contexts, and find that this encourages the emergence of robust behaviours that perform well across a suite of tasks. We demonstrate this principle for locomotion -- behaviours that are known for their sensitivity to the choice of reward. We train several simulated bodies on a diverse set of challenging terrains and obstacles, using a simple reward function based on forward progress. Using a novel scalable variant of policy gradient reinforcement learning, our agents learn to run, jump, crouch and turn as required by the environment without explicit reward-based guidance. A visual depiction of highlights of the learned behavior can be viewed following https://youtu.be/hx_bgoTF7bs .},761 archiveprefix = {arXiv},762 keywords = {Computer Science - Artificial Intelligence},763 file = {/Users/fracapuano/Zotero/storage/9DZ8XEVY/Heess et al. - 2017 - Emergence of Locomotion Behaviours in Rich Environments.pdf;/Users/fracapuano/Zotero/storage/JUB2Q3WH/1707.html}764}765 766@inproceedings{higgins2017beta,767 title = {Beta-Vae: {Learning} Basic Visual Concepts with a Constrained Variational Framework},768 booktitle = {International Conference on Learning Representations},769 author = {Higgins, Irina and Matthey, Loic and Pal, Arka and Burgess, Christopher and Glorot, Xavier and Botvinick, Matthew and Mohamed, Shakir and Lerchner, Alexander},770 year = {2017}771}772 773@misc{hoDenoisingDiffusionProbabilistic2020,774 title = {Denoising {Diffusion Probabilistic Models}},775 author = {Ho, Jonathan and Jain, Ajay and Abbeel, Pieter},776 year = {2020},777 month = dec,778 number = {arXiv:2006.11239},779 eprint = {2006.11239},780 primaryclass = {cs},781 publisher = {arXiv},782 doi = {10.48550/arXiv.2006.11239},783 urldate = {2025-09-03},784 abstract = {We present high quality image synthesis results using diffusion probabilistic models, a class of latent variable models inspired by considerations from nonequilibrium thermodynamics. Our best results are obtained by training on a weighted variational bound designed according to a novel connection between diffusion probabilistic models and denoising score matching with Langevin dynamics, and our models naturally admit a progressive lossy decompression scheme that can be interpreted as a generalization of autoregressive decoding. On the unconditional CIFAR10 dataset, we obtain an Inception score of 9.46 and a state-of-the-art FID score of 3.17. On 256x256 LSUN, we obtain sample quality similar to ProgressiveGAN. Our implementation is available at https://github.com/hojonathanho/diffusion},785 archiveprefix = {arXiv},786 keywords = {Computer Science - Machine Learning,Statistics - Machine Learning},787 file = {/Users/fracapuano/Zotero/storage/DE655AYQ/Ho et al. - 2020 - Denoising Diffusion Probabilistic Models.pdf;/Users/fracapuano/Zotero/storage/NVIS47ZH/2006.html}788}789 790@article{hwangboLearningAgileDynamic2019,791 title = {Learning Agile and Dynamic Motor Skills for Legged Robots},792 author = {Hwangbo, Jemin and Lee, Joonho and Dosovitskiy, Alexey and Bellicoso, Dario and Tsounis, Vassilios and Koltun, Vladlen and Hutter, Marco},793 year = {2019},794 month = jan,795 journal = {Science Robotics},796 volume = {4},797 number = {26},798 pages = {eaau5872},799 publisher = {American Association for the Advancement of Science},800 doi = {10.1126/scirobotics.aau5872},801 urldate = {2025-08-27},802 abstract = {Legged robots pose one of the greatest challenges in robotics. Dynamic and agile maneuvers of animals cannot be imitated by existing methods that are crafted by humans. A compelling alternative is reinforcement learning, which requires minimal craftsmanship and promotes the natural evolution of a control policy. However, so far, reinforcement learning research for legged robots is mainly limited to simulation, and only few and comparably simple examples have been deployed on real systems. The primary reason is that training with real robots, particularly with dynamically balancing systems, is complicated and expensive. In the present work, we introduce a method for training a neural network policy in simulation and transferring it to a state-of-the-art legged system, thereby leveraging fast, automated, and cost-effective data generation schemes. The approach is applied to the ANYmal robot, a sophisticated medium-dog--sized quadrupedal system. Using policies trained in simulation, the quadrupedal machine achieves locomotion skills that go beyond what had been achieved with prior methods: ANYmal is capable of precisely and energy-efficiently following high-level body velocity commands, running faster than before, and recovering from falling even in complex configurations.},803 file = {/Users/fracapuano/Zotero/storage/9V3X2F7R/Hwangbo et al. - 2019 - Learning agile and dynamic motor skills for legged robots.pdf}804}805 806@inproceedings{ImageNet_VSS09,807 title = {Construction and Analysis of a Large Scale Image Ontology},808 author = {Deng, J. and Li, K. and Do, M. and Su, H. and {Fei-Fei}, L.},809 year = {2009},810 publisher = {Vision Sciences Society}811}812 813@inproceedings{InstructBLIP,814 title = {InstructBLIP: {Towards} General-Purpose Vision-Language Models with Instruction Tuning},815 booktitle = {Thirty-Seventh Conference on Neural Information Processing Systems},816 author = {Dai, Wenliang and Li, Junnan and Li, Dongxu and Tiong, Anthony and Zhao, Junqi and Wang, Weisheng and Li, Boyang and Fung, Pascale and Hoi, Steven},817 year = {2023}818}819 820@misc{jangBCZZeroShotTask2022,821 title = {BC-Z: Zero-Shot Task Generalization with {Robotic Imitation Learning}},822 shorttitle = {BC-Z},823 author = {Jang, Eric and Irpan, Alex and Khansari, Mohi and Kappler, Daniel and Ebert, Frederik and Lynch, Corey and Levine, Sergey and Finn, Chelsea},824 year = {2022},825 month = feb,826 number = {arXiv:2202.02005},827 eprint = {2202.02005},828 primaryclass = {cs},829 publisher = {arXiv},830 doi = {10.48550/arXiv.2202.02005},831 urldate = {2025-09-01},832 abstract = {In this paper, we study the problem of enabling a vision-based robotic manipulation system to generalize to novel tasks, a long-standing challenge in robot learning. We approach the challenge from an imitation learning perspective, aiming to study how scaling and broadening the data collected can facilitate such generalization. To that end, we develop an interactive and flexible imitation learning system that can learn from both demonstrations and interventions and can be conditioned on different forms of information that convey the task, including pre-trained embeddings of natural language or videos of humans performing the task. When scaling data collection on a real robot to more than 100 distinct tasks, we find that this system can perform 24 unseen manipulation tasks with an average success rate of 44%, without any robot demonstrations for those tasks.},833 archiveprefix = {arXiv},834 keywords = {Computer Science - Machine Learning,Computer Science - Robotics},835 file = {/Users/fracapuano/Zotero/storage/YDG2WMDC/Jang et al. - 2022 - BC-Z Zero-Shot Task Generalization with Robotic Imitation Learning.pdf;/Users/fracapuano/Zotero/storage/ZZ47RG6V/2202.html}836}837 838@misc{jannerPlanningDiffusionFlexible2022,839 title = {Planning with {Diffusion} for {Flexible Behavior Synthesis}},840 author = {Janner, Michael and Du, Yilun and Tenenbaum, Joshua B. and Levine, Sergey},841 year = {2022},842 month = dec,843 number = {arXiv:2205.09991},844 eprint = {2205.09991},845 primaryclass = {cs},846 publisher = {arXiv},847 doi = {10.48550/arXiv.2205.09991},848 urldate = {2025-09-03},849 abstract = {Model-based reinforcement learning methods often use learning only for the purpose of estimating an approximate dynamics model, offloading the rest of the decision-making work to classical trajectory optimizers. While conceptually simple, this combination has a number of empirical shortcomings, suggesting that learned models may not be well-suited to standard trajectory optimization. In this paper, we consider what it would look like to fold as much of the trajectory optimization pipeline as possible into the modeling problem, such that sampling from the model and planning with it become nearly identical. The core of our technical approach lies in a diffusion probabilistic model that plans by iteratively denoising trajectories. We show how classifier-guided sampling and image inpainting can be reinterpreted as coherent planning strategies, explore the unusual and useful properties of diffusion-based planning methods, and demonstrate the effectiveness of our framework in control settings that emphasize long-horizon decision-making and test-time flexibility.},850 archiveprefix = {arXiv},851 keywords = {Computer Science - Artificial Intelligence,Computer Science - Machine Learning},852 file = {/Users/fracapuano/Zotero/storage/6S28T733/Janner et al. - 2022 - Planning with Diffusion for Flexible Behavior Synthesis.pdf;/Users/fracapuano/Zotero/storage/DRH9ZWCG/2205.html}853}854 855@misc{jiangMistral7B2023,856 title = {Mistral 7B},857 author = {Jiang, Albert Q. and Sablayrolles, Alexandre and Mensch, Arthur and Bamford, Chris and Chaplot, Devendra Singh and de las Casas, Diego and Bressand, Florian and Lengyel, Gianna and Lample, Guillaume and Saulnier, Lucile and Lavaud, L{\'e}lio Renard and Lachaux, Marie-Anne and Stock, Pierre and Scao, Teven Le and Lavril, Thibaut and Wang, Thomas and Lacroix, Timoth{\'e}e and Sayed, William El},858 year = {2023},859 month = oct,860 number = {arXiv:2310.06825},861 eprint = {2310.06825},862 primaryclass = {cs},863 publisher = {arXiv},864 doi = {10.48550/arXiv.2310.06825},865 urldate = {2025-09-09},866 abstract = {We introduce Mistral 7B v0.1, a 7-billion-parameter language model engineered for superior performance and efficiency. Mistral 7B outperforms Llama 2 13B across all evaluated benchmarks, and Llama 1 34B in reasoning, mathematics, and code generation. Our model leverages grouped-query attention (GQA) for faster inference, coupled with sliding window attention (SWA) to effectively handle sequences of arbitrary length with a reduced inference cost. We also provide a model fine-tuned to follow instructions, Mistral 7B -- Instruct, that surpasses the Llama 2 13B -- Chat model both on human and automated benchmarks. Our models are released under the Apache 2.0 license.},867 archiveprefix = {arXiv},868 keywords = {Computer Science - Artificial Intelligence,Computer Science - Computation and Language,Computer Science - Machine Learning},869 file = {/Users/fracapuano/Zotero/storage/JJX9Q8J9/Jiang et al. - 2023 - Mistral 7B.pdf;/Users/fracapuano/Zotero/storage/WTMQBRW3/2310.html}870}871 872@misc{jiDribbleBotDynamicLegged2023,873 title = {DribbleBot: {Dynamic Legged Manipulation} in the {Wild}},874 shorttitle = {DribbleBot},875 author = {Ji, Yandong and Margolis, Gabriel B. and Agrawal, Pulkit},876 year = {2023},877 month = apr,878 number = {arXiv:2304.01159},879 eprint = {2304.01159},880 primaryclass = {cs},881 publisher = {arXiv},882 doi = {10.48550/arXiv.2304.01159},883 urldate = {2025-08-26},884 abstract = {DribbleBot (Dexterous Ball Manipulation with a Legged Robot) is a legged robotic system that can dribble a soccer ball under the same real-world conditions as humans (i.e., in-the-wild). We adopt the paradigm of training policies in simulation using reinforcement learning and transferring them into the real world. We overcome critical challenges of accounting for variable ball motion dynamics on different terrains and perceiving the ball using body-mounted cameras under the constraints of onboard computing. Our results provide evidence that current quadruped platforms are well-suited for studying dynamic whole-body control problems involving simultaneous locomotion and manipulation directly from sensory observations.},885 archiveprefix = {arXiv},886 keywords = {Computer Science - Artificial Intelligence,Computer Science - Machine Learning,Computer Science - Robotics},887 file = {/Users/fracapuano/Zotero/storage/ABSRE4C4/Ji et al. - 2023 - DribbleBot Dynamic Legged Manipulation in the Wild.pdf;/Users/fracapuano/Zotero/storage/ADI4QNCY/2304.html}888}889 890@misc{kakaobrain2022coyo700m,891 title = {COYO-700M: Image-text Pair Dataset},892 author = {Byeon, Minwoo and Park, Beomhee and Kim, Haecheon and Lee, Sungjun and Baek, Woonhyuk and Kim, Saehoon},893 year = {2022}894}895 896@misc{kaplanScalingLawsNeural2020,897 title = {Scaling {Laws} for {Neural Language Models}},898 author = {Kaplan, Jared and McCandlish, Sam and Henighan, Tom and Brown, Tom B. and Chess, Benjamin and Child, Rewon and Gray, Scott and Radford, Alec and Wu, Jeffrey and Amodei, Dario},899 year = {2020},900 month = jan,901 number = {arXiv:2001.08361},902 eprint = {2001.08361},903 primaryclass = {cs},904 publisher = {arXiv},905 doi = {10.48550/arXiv.2001.08361},906 urldate = {2025-09-07},907 abstract = {We study empirical scaling laws for language model performance on the cross-entropy loss. The loss scales as a power-law with model size, dataset size, and the amount of compute used for training, with some trends spanning more than seven orders of magnitude. Other architectural details such as network width or depth have minimal effects within a wide range. Simple equations govern the dependence of overfitting on model/dataset size and the dependence of training speed on model size. These relationships allow us to determine the optimal allocation of a fixed compute budget. Larger models are significantly more sample-efficient, such that optimally compute-efficient training involves training very large models on a relatively modest amount of data and stopping significantly before convergence.},908 archiveprefix = {arXiv},909 keywords = {Computer Science - Machine Learning,Statistics - Machine Learning},910 file = {/Users/fracapuano/Zotero/storage/MI5AGWBH/Kaplan et al. - 2020 - Scaling Laws for Neural Language Models.pdf;/Users/fracapuano/Zotero/storage/SBZT8DDY/2001.html}911}912 913@misc{keGraspingChopsticksCombating2020,914 title = {Grasping with {Chopsticks}: {Combating Covariate Shift} in Model-free Imitation Learning for {Fine Manipulation}},915 shorttitle = {Grasping with {Chopsticks}},916 author = {Ke, Liyiming and Wang, Jingqiang and Bhattacharjee, Tapomayukh and Boots, Byron and Srinivasa, Siddhartha},917 year = {2020},918 month = nov,919 number = {arXiv:2011.06719},920 eprint = {2011.06719},921 primaryclass = {cs},922 publisher = {arXiv},923 doi = {10.48550/arXiv.2011.06719},924 urldate = {2025-09-01},925 abstract = {Billions of people use chopsticks, a simple yet versatile tool, for fine manipulation of everyday objects. The small, curved, and slippery tips of chopsticks pose a challenge for picking up small objects, making them a suitably complex test case. This paper leverages human demonstrations to develop an autonomous chopsticks-equipped robotic manipulator. Due to the lack of accurate models for fine manipulation, we explore model-free imitation learning, which traditionally suffers from the covariate shift phenomenon that causes poor generalization. We propose two approaches to reduce covariate shift, neither of which requires access to an interactive expert or a model, unlike previous approaches. First, we alleviate single-step prediction errors by applying an invariant operator to increase the data support at critical steps for grasping. Second, we generate synthetic corrective labels by adding bounded noise and combining parametric and non-parametric methods to prevent error accumulation. We demonstrate our methods on a real chopstick-equipped robot that we built, and observe the agent's success rate increase from 37.3% to 80%, which is comparable to the human expert performance of 82.6%.},926 archiveprefix = {arXiv},927 keywords = {Computer Science - Machine Learning,Computer Science - Robotics},928 file = {/Users/fracapuano/Zotero/storage/ZUPECLSW/Ke et al. - 2020 - Grasping with Chopsticks Combating Covariate Shift in Model-free Imitation Learning for Fine Manipu.pdf;/Users/fracapuano/Zotero/storage/X7PX638S/2011.html}929}930 931@article{khatibRealTimeObstancleAvoidance1986,932 title = {Real-{Time Obstancle Avoidance} for {Manipulators} and {Mobile Robots}},933 author = {Khatib, Oussama},934 year = {1986},935 journal = {The International Journal of Robotics Research},936 volume = {5}937}938 939@misc{khazatskyDROIDLargeScaleInTheWild2025,940 title = {DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset},941 shorttitle = {DROID},942 author = {Khazatsky, Alexander and Pertsch, Karl and Nair, Suraj and Balakrishna, Ashwin and Dasari, Sudeep and Karamcheti, Siddharth and Nasiriany, Soroush and Srirama, Mohan Kumar and Chen, Lawrence Yunliang and Ellis, Kirsty and Fagan, Peter David and Hejna, Joey and Itkina, Masha and Lepert, Marion and Ma, Yecheng Jason and Miller, Patrick Tree and Wu, Jimmy and Belkhale, Suneel and Dass, Shivin and Ha, Huy and Jain, Arhan and Lee, Abraham and Lee, Youngwoon and Memmel, Marius and Park, Sungjae and Radosavovic, Ilija and Wang, Kaiyuan and Zhan, Albert and Black, Kevin and Chi, Cheng and Hatch, Kyle Beltran and Lin, Shan and Lu, Jingpei and Mercat, Jean and Rehman, Abdul and Sanketi, Pannag R. and Sharma, Archit and Simpson, Cody and Vuong, Quan and Walke, Homer Rich and Wulfe, Blake and Xiao, Ted and Yang, Jonathan Heewon and Yavary, Arefeh and Zhao, Tony Z. and Agia, Christopher and Baijal, Rohan and Castro, Mateo Guaman and Chen, Daphne and Chen, Qiuyu and Chung, Trinity and Drake, Jaimyn and Foster, Ethan Paul and Gao, Jensen and Guizilini, Vitor and Herrera, David Antonio and Heo, Minho and Hsu, Kyle and Hu, Jiaheng and Irshad, Muhammad Zubair and Jackson, Donovon and Le, Charlotte and Li, Yunshuang and Lin, Kevin and Lin, Roy and Ma, Zehan and Maddukuri, Abhiram and Mirchandani, Suvir and Morton, Daniel and Nguyen, Tony and O'Neill, Abigail and Scalise, Rosario and Seale, Derick and Son, Victor and Tian, Stephen and Tran, Emi and Wang, Andrew E. and Wu, Yilin and Xie, Annie and Yang, Jingyun and Yin, Patrick and Zhang, Yunchu and Bastani, Osbert and Berseth, Glen and Bohg, Jeannette and Goldberg, Ken and Gupta, Abhinav and Gupta, Abhishek and Jayaraman, Dinesh and Lim, Joseph J. and Malik, Jitendra and {Mart{\'i}n-Mart{\'i}n}, Roberto and Ramamoorthy, Subramanian and Sadigh, Dorsa and Song, Shuran and Wu, Jiajun and Yip, Michael C. and Zhu, Yuke and Kollar, Thomas and Levine, Sergey and Finn, Chelsea},943 year = {2025},944 month = apr,945 number = {arXiv:2403.12945},946 eprint = {2403.12945},947 primaryclass = {cs},948 publisher = {arXiv},949 doi = {10.48550/arXiv.2403.12945},950 urldate = {2025-09-08},951 abstract = {The creation of large, diverse, high-quality robot manipulation datasets is an important stepping stone on the path toward more capable and robust robotic manipulation policies. However, creating such datasets is challenging: collecting robot manipulation data in diverse environments poses logistical and safety challenges and requires substantial investments in hardware and human labour. As a result, even the most general robot manipulation policies today are mostly trained on data collected in a small number of environments with limited scene and task diversity. In this work, we introduce DROID (Distributed Robot Interaction Dataset), a diverse robot manipulation dataset with 76k demonstration trajectories or 350 hours of interaction data, collected across 564 scenes and 84 tasks by 50 data collectors in North America, Asia, and Europe over the course of 12 months. We demonstrate that training with DROID leads to policies with higher performance and improved generalization ability. We open source the full dataset, policy learning code, and a detailed guide for reproducing our robot hardware setup.},952 archiveprefix = {arXiv},953 keywords = {Computer Science - Robotics},954 file = {/Users/fracapuano/Zotero/storage/XZ5Y4HZS/Khazatsky et al. - 2025 - DROID A Large-Scale In-The-Wild Robot Manipulation Dataset.pdf;/Users/fracapuano/Zotero/storage/N2Z72XLK/2403.html}955}956 957@misc{kimOpenVLAOpenSourceVisionLanguageAction2024,958 title = {OpenVLA: An Open-Source Vision-Language-Action Model},959 shorttitle = {OpenVLA},960 author = {Kim, Moo Jin and Pertsch, Karl and Karamcheti, Siddharth and Xiao, Ted and Balakrishna, Ashwin and Nair, Suraj and Rafailov, Rafael and Foster, Ethan and Lam, Grace and Sanketi, Pannag and Vuong, Quan and Kollar, Thomas and Burchfiel, Benjamin and Tedrake, Russ and Sadigh, Dorsa and Levine, Sergey and Liang, Percy and Finn, Chelsea},961 year = {2024},962 month = sep,963 number = {arXiv:2406.09246},964 eprint = {2406.09246},965 primaryclass = {cs},966 publisher = {arXiv},967 doi = {10.48550/arXiv.2406.09246},968 urldate = {2025-09-08},969 abstract = {Large policies pretrained on a combination of Internet-scale vision-language data and diverse robot demonstrations have the potential to change how we teach robots new skills: rather than training new behaviors from scratch, we can fine-tune such vision-language-action (VLA) models to obtain robust, generalizable policies for visuomotor control. Yet, widespread adoption of VLAs for robotics has been challenging as 1) existing VLAs are largely closed and inaccessible to the public, and 2) prior work fails to explore methods for efficiently fine-tuning VLAs for new tasks, a key component for adoption. Addressing these challenges, we introduce OpenVLA, a 7B-parameter open-source VLA trained on a diverse collection of 970k real-world robot demonstrations. OpenVLA builds on a Llama 2 language model combined with a visual encoder that fuses pretrained features from DINOv2 and SigLIP. As a product of the added data diversity and new model components, OpenVLA demonstrates strong results for generalist manipulation, outperforming closed models such as RT-2-X (55B) by 16.5% in absolute task success rate across 29 tasks and multiple robot embodiments, with 7x fewer parameters. We further show that we can effectively fine-tune OpenVLA for new settings, with especially strong generalization results in multi-task environments involving multiple objects and strong language grounding abilities, and outperform expressive from-scratch imitation learning methods such as Diffusion Policy by 20.4%. We also explore compute efficiency; as a separate contribution, we show that OpenVLA can be fine-tuned on consumer GPUs via modern low-rank adaptation methods and served efficiently via quantization without a hit to downstream success rate. Finally, we release model checkpoints, fine-tuning notebooks, and our PyTorch codebase with built-in support for training VLAs at scale on Open X-Embodiment datasets.},970 archiveprefix = {arXiv},971 keywords = {Computer Science - Machine Learning,Computer Science - Robotics},972 file = {/Users/fracapuano/Zotero/storage/XR2SX8WG/Kim et al. - 2024 - OpenVLA An Open-Source Vision-Language-Action Model.pdf;/Users/fracapuano/Zotero/storage/63Q96WRV/2406.html}973}974 975@misc{kingmaAutoEncodingVariationalBayes2022,976 title = {Auto-{Encoding Variational Bayes}},977 author = {Kingma, Diederik P. and Welling, Max},978 year = {2022},979 month = dec,980 number = {arXiv:1312.6114},981 eprint = {1312.6114},982 primaryclass = {stat},983 publisher = {arXiv},984 doi = {10.48550/arXiv.1312.6114},985 urldate = {2025-09-02},986 abstract = {How can we perform efficient inference and learning in directed probabilistic models, in the presence of continuous latent variables with intractable posterior distributions, and large datasets? We introduce a stochastic variational inference and learning algorithm that scales to large datasets and, under some mild differentiability conditions, even works in the intractable case. Our contributions are two-fold. First, we show that a reparameterization of the variational lower bound yields a lower bound estimator that can be straightforwardly optimized using standard stochastic gradient methods. Second, we show that for i.i.d. datasets with continuous latent variables per datapoint, posterior inference can be made especially efficient by fitting an approximate inference model (also called a recognition model) to the intractable posterior using the proposed lower bound estimator. Theoretical advantages are reflected in experimental results.},987 archiveprefix = {arXiv},988 keywords = {Computer Science - Machine Learning,Statistics - Machine Learning},989 file = {/Users/fracapuano/Zotero/storage/IT7VNQ4U/Kingma and Welling - 2022 - Auto-Encoding Variational Bayes.pdf;/Users/fracapuano/Zotero/storage/HQT22HP5/1312.html}990}991 992@misc{knightStandardOpenSO100,993 title = {Standard Open SO-100 & SO-101 Arms},994 author = {Knight, Rob and Kooijmans, Pepijn and Wolf, Thomas and Alibert, Simon and Aractingi, Michel and Aubakirova, Dana and Zouitine, Adil and Martino, Russi and Palma, Steven and Pascal, Caroline and Cadene, Remi}995}996 997@article{koberReinforcementLearningRobotics,998 title = {Reinforcement {Learning} in {Robotics}: {A Survey}},999 author = {Kober, Jens and Bagnell, J Andrew and Peters, Jan},1000 langid = {english},1001 file = {/Users/fracapuano/Zotero/storage/72PRHGKL/Kober et al. - Reinforcement Learning in Robotics A Survey.pdf}1002}1003 1004@inproceedings{kong2024audioflam,1005 title = {Audio Flamingo: A Novel Audio Language Model with Few-Shot Learning and Dialogue Abilities},1006 booktitle = {International Conference on Machine Learning},1007 author = {Kong, Zhifeng and Goel, Arushi and Badlani, Rohan and Ping, Wei and Valle, Rafael and Catanzaro, Bryan},1008 year = {2024},1009 pages = {25125--25148},1010 publisher = {PMLR}1011}1012 1013@misc{kumarRMARapidMotor2021,1014 title = {RMA: {Rapid Motor Adaptation} for {Legged Robots}},1015 shorttitle = {RMA},1016 author = {Kumar, Ashish and Fu, Zipeng and Pathak, Deepak and Malik, Jitendra},1017 year = {2021},1018 month = jul,1019 number = {arXiv:2107.04034},1020 eprint = {2107.04034},1021 primaryclass = {cs},1022 publisher = {arXiv},1023 doi = {10.48550/arXiv.2107.04034},1024 urldate = {2025-08-27},1025 abstract = {Successful real-world deployment of legged robots would require them to adapt in real-time to unseen scenarios like changing terrains, changing payloads, wear and tear. This paper presents Rapid Motor Adaptation (RMA) algorithm to solve this problem of real-time online adaptation in quadruped robots. RMA consists of two components: a base policy and an adaptation module. The combination of these components enables the robot to adapt to novel situations in fractions of a second. RMA is trained completely in simulation without using any domain knowledge like reference trajectories or predefined foot trajectory generators and is deployed on the A1 robot without any fine-tuning. We train RMA on a varied terrain generator using bioenergetics-inspired rewards and deploy it on a variety of difficult terrains including rocky, slippery, deformable surfaces in environments with grass, long vegetation, concrete, pebbles, stairs, sand, etc. RMA shows state-of-the-art performance across diverse real-world as well as simulation experiments. Video results at https://ashish-kmr.github.io/rma-legged-robots/},1026 archiveprefix = {arXiv},1027 keywords = {Computer Science - Artificial Intelligence,Computer Science - Computer Vision and Pattern Recognition,Computer Science - Machine Learning,Computer Science - Robotics},1028 file = {/Users/fracapuano/Zotero/storage/TMYICHS6/Kumar et al. - 2021 - RMA Rapid Motor Adaptation for Legged Robots.pdf;/Users/fracapuano/Zotero/storage/TFY2EU8I/2107.html}1029}1030 1031@misc{laiActionChunkingConditional2025,1032 title = {Action Chunking as Conditional Policy Compression},1033 author = {Lai, Lucy and Huang, Ann and Gershman, Samuel},1034 year = {2025},1035 month = jun,1036 publisher = {OSF},1037 doi = {10.31234/osf.io/z8yrv_v2},1038 urldate = {2025-09-02},1039 abstract = {Many skills in our everyday lives are learned by sequencing actions towards a desired goal. The action sequence can become a ``chunk'' when individual actions are grouped together and executed as one unit, making them more efficient to store and execute. While chunking has been studied extensively across various domains, a puzzle remains as to why and under what conditions action chunking occurs. To tackle these questions, we develop a model of conditional policy compression---the reduction in cognitive cost by conditioning on an additional source of information---to explain the origin of chunking. We argue that chunking is a result of optimizing the trade-off between reward and conditional policy complexity. Chunking compresses policies when there is temporal structure in the environment that can be leveraged for action selection, reducing the amount of memory necessary to encode the policy. We experimentally confirm our model's predictions, showing that chunking reduces conditional policy complexity and reaction times. Chunking also increases with working memory load, consistent with the hypothesis that the degree of policy compression scales with the scarcity of cognitive resources. Finally, chunking also reduces overall working memory load, freeing cognitive resources for the benefit of other, not-chunked information.},1040 archiveprefix = {OSF},1041 langid = {american},1042 keywords = {action selection,chunking,habits,reinforcement learning,resource-rationality,working memory}1043}1044 1045@article{laiActionChunkingConditional2025a,1046 title = {Action Chunking as Conditional Policy Compression},1047 author = {Lai, Lucy and Huang, Ann Z. X. and Gershman, Samuel J.},1048 year = {2025},1049 month = nov,1050 journal = {Cognition},1051 volume = {264},1052 pages = {106201},1053 issn = {1873-7838},1054 doi = {10.1016/j.cognition.2025.106201},1055 abstract = {Many skills in our everyday lives are learned by sequencing actions towards a desired goal. The action sequence can become a "chunk" when individual actions are grouped together and executed as one unit, making them more efficient to store and execute. While chunking has been studied extensively across various domains, a puzzle remains as to why and under what conditions action chunking occurs. To tackle these questions, we develop a model of conditional policy compression-the reduction in cognitive cost by conditioning on an additional source of information-to explain the origin of chunking. We argue that chunking is a result of optimizing the trade-off between reward and conditional policy complexity. Chunking compresses policies when there is temporal structure in the environment that can be leveraged for action selection, reducing the amount of memory necessary to encode the policy. We experimentally confirm our model's predictions, showing that chunking reduces conditional policy complexity and reaction times. Chunking also increases with working memory load, consistent with the hypothesis that the degree of policy compression scales with the scarcity of cognitive resources. Finally, chunking also reduces overall working memory load, freeing cognitive resources for the benefit of other, not-chunked information.},1056 langid = {english},1057 pmid = {40602234},1058 keywords = {Action selection,Adult,Chunking,Cognition,Decision making,Female,Humans,Information bottleneck,Male,Memory Short-Term,Models Psychological,Psychomotor Performance,Reaction Time,Reinforcement learning,Resource rationality,Reward,Young Adult}1059}1060 1061@article{LAION-COCO,1062 title = {Laion Coco: 600m Synthetic Captions from Laion2b-En},1063 author = {Schuhmann, C and K{\"o}pf, A and Vencu, R and Coombes, T and Beaumont, R},1064 year = {2022},1065 journal = {URL https://laion.ai/blog/laion-coco}1066}1067 1068@misc{laurenconWhatMattersWhen2024,1069 title = {What Matters When Building Vision-Language Models?},1070 author = {Lauren{\c c}on, Hugo and Tronchon, L{\'e}o and Cord, Matthieu and Sanh, Victor},1071 year = {2024},1072 month = may,1073 number = {arXiv:2405.02246},1074 eprint = {2405.02246},1075 primaryclass = {cs},1076 publisher = {arXiv},1077 doi = {10.48550/arXiv.2405.02246},1078 urldate = {2025-09-09},1079 abstract = {The growing interest in vision-language models (VLMs) has been driven by improvements in large language models and vision transformers. Despite the abundance of literature on this subject, we observe that critical decisions regarding the design of VLMs are often not justified. We argue that these unsupported decisions impede progress in the field by making it difficult to identify which choices improve model performance. To address this issue, we conduct extensive experiments around pre-trained models, architecture choice, data, and training methods. Our consolidation of findings includes the development of Idefics2, an efficient foundational VLM of 8 billion parameters. Idefics2 achieves state-of-the-art performance within its size category across various multimodal benchmarks, and is often on par with models four times its size. We release the model (base, instructed, and chat) along with the datasets created for its training.},1080 archiveprefix = {arXiv},1081 keywords = {Computer Science - Artificial Intelligence,Computer Science - Computer Vision and Pattern Recognition},1082 file = {/Users/fracapuano/Zotero/storage/8H6NRPU7/Laurençon et al. - 2024 - What matters when building vision-language models.pdf;/Users/fracapuano/Zotero/storage/H3NETYXA/2405.html}1083}1084 1085@misc{leeBehaviorGenerationLatent2024,1086 title = {Behavior {Generation} with {Latent Actions}},1087 author = {Lee, Seungjae and Wang, Yibin and Etukuru, Haritheja and Kim, H. Jin and Shafiullah, Nur Muhammad Mahi and Pinto, Lerrel},1088 year = {2024},1089 month = jun,1090 number = {arXiv:2403.03181},1091 eprint = {2403.03181},1092 primaryclass = {cs},1093 publisher = {arXiv},1094 doi = {10.48550/arXiv.2403.03181},1095 urldate = {2025-08-28},1096 abstract = {Generative modeling of complex behaviors from labeled datasets has been a longstanding problem in decision making. Unlike language or image generation, decision making requires modeling actions - continuous-valued vectors that are multimodal in their distribution, potentially drawn from uncurated sources, where generation errors can compound in sequential prediction. A recent class of models called Behavior Transformers (BeT) addresses this by discretizing actions using k-means clustering to capture different modes. However, k-means struggles to scale for high-dimensional action spaces or long sequences, and lacks gradient information, and thus BeT suffers in modeling long-range actions. In this work, we present Vector-Quantized Behavior Transformer (VQ-BeT), a versatile model for behavior generation that handles multimodal action prediction, conditional generation, and partial observations. VQ-BeT augments BeT by tokenizing continuous actions with a hierarchical vector quantization module. Across seven environments including simulated manipulation, autonomous driving, and robotics, VQ-BeT improves on state-of-the-art models such as BeT and Diffusion Policies. Importantly, we demonstrate VQ-BeT's improved ability to capture behavior modes while accelerating inference speed 5x over Diffusion Policies. Videos and code can be found https://sjlee.cc/vq-bet},1097 archiveprefix = {arXiv},1098 keywords = {Computer Science - Artificial Intelligence,Computer Science - Machine Learning,Computer Science - Robotics},1099 file = {/Users/fracapuano/Zotero/storage/IA93ENCH/Lee et al. - 2024 - Behavior Generation with Latent Actions.pdf;/Users/fracapuano/Zotero/storage/KBVF7GQL/2403.html}1100}1101 1102@article{leeLearningQuadrupedalLocomotion2020,1103 title = {Learning {Quadrupedal Locomotion} over {Challenging Terrain}},1104 author = {Lee, Joonho and Hwangbo, Jemin and Wellhausen, Lorenz and Koltun, Vladlen and Hutter, Marco},1105 year = {2020},1106 month = oct,1107 journal = {Science Robotics},1108 volume = {5},1109 number = {47},1110 eprint = {2010.11251},1111 primaryclass = {cs},1112 pages = {eabc5986},1113 issn = {2470-9476},1114 doi = {10.1126/scirobotics.abc5986},1115 urldate = {2025-08-26},1116 abstract = {Some of the most challenging environments on our planet are accessible to quadrupedal animals but remain out of reach for autonomous machines. Legged locomotion can dramatically expand the operational domains of robotics. However, conventional controllers for legged locomotion are based on elaborate state machines that explicitly trigger the execution of motion primitives and reflexes. These designs have escalated in complexity while falling short of the generality and robustness of animal locomotion. Here we present a radically robust controller for legged locomotion in challenging natural environments. We present a novel solution to incorporating proprioceptive feedback in locomotion control and demonstrate remarkable zero-shot generalization from simulation to natural environments. The controller is trained by reinforcement learning in simulation. It is based on a neural network that acts on a stream of proprioceptive signals. The trained controller has taken two generations of quadrupedal ANYmal robots to a variety of natural environments that are beyond the reach of prior published work in legged locomotion. The controller retains its robustness under conditions that have never been encountered during training: deformable terrain such as mud and snow, dynamic footholds such as rubble, and overground impediments such as thick vegetation and gushing water. The presented work opens new frontiers for robotics and indicates that radical robustness in natural environments can be achieved by training in much simpler domains.},1117 archiveprefix = {arXiv},1118 keywords = {Computer Science - Machine Learning,Computer Science - Robotics,Computer Science - Systems and Control,Electrical Engineering and Systems Science - Systems and Control},1119 file = {/Users/fracapuano/Zotero/storage/8B9EF2CE/Lee et al. - 2020 - Learning Quadrupedal Locomotion over Challenging Terrain.pdf}1120}1121 1122@misc{lillicrapContinuousControlDeep2019,1123 title = {Continuous Control with Deep Reinforcement Learning},1124 author = {Lillicrap, Timothy P. and Hunt, Jonathan J. and Pritzel, Alexander and Heess, Nicolas and Erez, Tom and Tassa, Yuval and Silver, David and Wierstra, Daan},1125 year = {2019},1126 month = jul,1127 number = {arXiv:1509.02971},1128 eprint = {1509.02971},1129 primaryclass = {cs},1130 publisher = {arXiv},1131 doi = {10.48550/arXiv.1509.02971},1132 urldate = {2025-08-31},1133 abstract = {We adapt the ideas underlying the success of Deep Q-Learning to the continuous action domain. We present an actor-critic, model-free algorithm based on the deterministic policy gradient that can operate over continuous action spaces. Using the same learning algorithm, network architecture and hyper-parameters, our algorithm robustly solves more than 20 simulated physics tasks, including classic problems such as cartpole swing-up, dexterous manipulation, legged locomotion and car driving. Our algorithm is able to find policies whose performance is competitive with those found by a planning algorithm with full access to the dynamics of the domain and its derivatives. We further demonstrate that for many of the tasks the algorithm can learn policies end-to-end: directly from raw pixel inputs.},1134 archiveprefix = {arXiv},1135 keywords = {Computer Science - Machine Learning,Statistics - Machine Learning},1136 file = {/Users/fracapuano/Zotero/storage/2VN6TMVK/Lillicrap et al. - 2019 - Continuous control with deep reinforcement learning.pdf;/Users/fracapuano/Zotero/storage/4FQ4W5VE/1509.html}1137}1138 1139@misc{lillicrapContinuousControlDeep2019a,1140 title = {Continuous Control with Deep Reinforcement Learning},1141 author = {Lillicrap, Timothy P. and Hunt, Jonathan J. and Pritzel, Alexander and Heess, Nicolas and Erez, Tom and Tassa, Yuval and Silver, David and Wierstra, Daan},1142 year = {2019},1143 month = jul,1144 number = {arXiv:1509.02971},1145 eprint = {1509.02971},1146 primaryclass = {cs},1147 publisher = {arXiv},1148 doi = {10.48550/arXiv.1509.02971},1149 urldate = {2025-08-31},1150 abstract = {We adapt the ideas underlying the success of Deep Q-Learning to the continuous action domain. We present an actor-critic, model-free algorithm based on the deterministic policy gradient that can operate over continuous action spaces. Using the same learning algorithm, network architecture and hyper-parameters, our algorithm robustly solves more than 20 simulated physics tasks, including classic problems such as cartpole swing-up, dexterous manipulation, legged locomotion and car driving. Our algorithm is able to find policies whose performance is competitive with those found by a planning algorithm with full access to the dynamics of the domain and its derivatives. We further demonstrate that for many of the tasks the algorithm can learn policies end-to-end: directly from raw pixel inputs.},1151 archiveprefix = {arXiv},1152 keywords = {Computer Science - Machine Learning,Statistics - Machine Learning},1153 file = {/Users/fracapuano/Zotero/storage/HYMPB9F5/Lillicrap et al. - 2019 - Continuous control with deep reinforcement learning.pdf;/Users/fracapuano/Zotero/storage/EKCXMJNQ/1509.html}1154}1155 1156@misc{linVILAPretrainingVisual2024,1157 title = {VILA: On Pre-training for {Visual Language Models}},1158 shorttitle = {VILA},1159 author = {Lin, Ji and Yin, Hongxu and Ping, Wei and Lu, Yao and Molchanov, Pavlo and Tao, Andrew and Mao, Huizi and Kautz, Jan and Shoeybi, Mohammad and Han, Song},1160 year = {2024},1161 month = may,1162 number = {arXiv:2312.07533},1163 eprint = {2312.07533},1164 primaryclass = {cs},1165 publisher = {arXiv},1166 doi = {10.48550/arXiv.2312.07533},1167 urldate = {2025-09-09},1168 abstract = {Visual language models (VLMs) rapidly progressed with the recent success of large language models. There have been growing efforts on visual instruction tuning to extend the LLM with visual inputs, but lacks an in-depth study of the visual language pre-training process, where the model learns to perform joint modeling on both modalities. In this work, we examine the design options for VLM pre-training by augmenting LLM towards VLM through step-by-step controllable comparisons. We introduce three main findings: (1) freezing LLMs during pre-training can achieve decent zero-shot performance, but lack in-context learning capability, which requires unfreezing the LLM; (2) interleaved pre-training data is beneficial whereas image-text pairs alone are not optimal; (3) re-blending text-only instruction data to image-text data during instruction fine-tuning not only remedies the degradation of text-only tasks, but also boosts VLM task accuracy. With an enhanced pre-training recipe we build VILA, a Visual Language model family that consistently outperforms the state-of-the-art models, e.g., LLaVA-1.5, across main benchmarks without bells and whistles. Multi-modal pre-training also helps unveil appealing properties of VILA, including multi-image reasoning, enhanced in-context learning, and better world knowledge.},1169 archiveprefix = {arXiv},1170 keywords = {Computer Science - Computer Vision and Pattern Recognition},1171 file = {/Users/fracapuano/Zotero/storage/DNA6AFRL/Lin et al. - 2024 - VILA On Pre-training for Visual Language Models.pdf;/Users/fracapuano/Zotero/storage/K32IJ2A3/2312.html}1172}1173 1174@misc{lipmanFlowMatchingGenerative2023,1175 title = {Flow {Matching} for {Generative Modeling}},1176 author = {Lipman, Yaron and Chen, Ricky T. Q. and {Ben-Hamu}, Heli and Nickel, Maximilian and Le, Matt},1177 year = {2023},1178 month = feb,1179 number = {arXiv:2210.02747},1180 eprint = {2210.02747},1181 primaryclass = {cs},1182 publisher = {arXiv},1183 doi = {10.48550/arXiv.2210.02747},1184 urldate = {2025-09-07},1185 abstract = {We introduce a new paradigm for generative modeling built on Continuous Normalizing Flows (CNFs), allowing us to train CNFs at unprecedented scale. Specifically, we present the notion of Flow Matching (FM), a simulation-free approach for training CNFs based on regressing vector fields of fixed conditional probability paths. Flow Matching is compatible with a general family of Gaussian probability paths for transforming between noise and data samples -- which subsumes existing diffusion paths as specific instances. Interestingly, we find that employing FM with diffusion paths results in a more robust and stable alternative for training diffusion models. Furthermore, Flow Matching opens the door to training CNFs with other, non-diffusion probability paths. An instance of particular interest is using Optimal Transport (OT) displacement interpolation to define the conditional probability paths. These paths are more efficient than diffusion paths, provide faster training and sampling, and result in better generalization. Training CNFs using Flow Matching on ImageNet leads to consistently better performance than alternative diffusion-based methods in terms of both likelihood and sample quality, and allows fast and reliable sample generation using off-the-shelf numerical ODE solvers.},1186 archiveprefix = {arXiv},1187 keywords = {Computer Science - Artificial Intelligence,Computer Science - Machine Learning,Statistics - Machine Learning},1188 file = {/Users/fracapuano/Zotero/storage/YFZTRGJ3/Lipman et al. - 2023 - Flow Matching for Generative Modeling.pdf;/Users/fracapuano/Zotero/storage/QUKPDHWR/2210.html}1189}1190 1191@misc{lipmanFlowMatchingGuide2024,1192 title = {Flow {Matching Guide} and {Code}},1193 author = {Lipman, Yaron and Havasi, Marton and Holderrieth, Peter and Shaul, Neta and Le, Matt and Karrer, Brian and Chen, Ricky T. Q. and {Lopez-Paz}, David and {Ben-Hamu}, Heli and Gat, Itai},1194 year = {2024},1195 month = dec,1196 number = {arXiv:2412.06264},1197 eprint = {2412.06264},1198 primaryclass = {cs},1199 publisher = {arXiv},1200 doi = {10.48550/arXiv.2412.06264},