Cache-SCA/Gr00t_n1.5-IsaacLab-SO101-Multi_Task-30fps_8epoch

1---2library_name: lerobot3pipeline_tag: robotics4base_model: nvidia/GR00T-N1.5-3B5base_model_relation: finetune6datasets:7- CoRL2026-CSI/Isaaclab-so101_11task_baseCaP_3300epi8license: other9license_name: nvidia-license10license_link: https://developer.download.nvidia.com/licenses/NVIDIA-OneWay-Noncommercial-License-22Mar2022.pdf11tags:12- robotics13- lerobot14- groot15- gr00t-n1.516- so10117- imitation-learning18- flow-matching19- safetensors20---21 22# CoRL2026-CSI/IsaacLab-so101-multi-gr00t23 24This is a LeRobot GR00T N1.5 policy fine-tuned from25[`nvidia/GR00T-N1.5-3B`](https://huggingface.co/nvidia/GR00T-N1.5-3B) on26[`CoRL2026-CSI/Isaaclab-so101_11task_baseCaP_3300epi`](https://huggingface.co/datasets/CoRL2026-CSI/Isaaclab-so101_11task_baseCaP_3300epi).27 28The model is intended for SO-101 style manipulation experiments using RGB observations,29robot proprioception, language instructions, and continuous action chunks. It is uploaded30as a LeRobot policy checkpoint and should be loaded through the matching LeRobot GR00T31implementation used for training.32 33## Model Details34 35- **Policy type:** GR00T N1.536- **Base model:** `nvidia/GR00T-N1.5-3B`37- **Tokenizer assets:** `lerobot/eagle2hg-processor-groot-n1p5`38- **Embodiment tag:** `new_embodiment`39- **Observation steps:** `1`40- **Action chunk size:** `16`41- **Action steps:** `16`42- **Max state/action dims:** `64` / `32`43 44## Fine-Tuning Setup45 46- **Training steps:** `110000`47- **Approx. epochs:** `7.99`48- **Final training samples:** `28160000`49- **Final training loss:** `0.010196`50- **Runtime:** `76.72 hours`51- **Micro batch size:** `64`52- **Gradient accumulation steps:** `2`53- **Effective batch size:** `256`54- **Optimizer:** `adamw`, lr `0.0001`, weight decay `1e-05`55- **Scheduler:** `cosine_decay_with_warmup`56- **Mixed precision:** bf16 policy weights/config enabled: `true`57- **VLM/LLM backbone fine-tuned:** `false`58- **Vision tower fine-tuned:** `false`59- **Action projector fine-tuned:** `true`60- **Action DiT fine-tuned:** `true`61 62The GR00T vision-language backbone was frozen for this run. The action head projector63and flow-matching diffusion transformer were fine-tuned.64 65## Inputs66 67- `observation.ee_pos.robot_xyzrpy`: `STATE`, shape `[6]`68- `observation.gripper_binary`: `STATE`, shape `[1]`69- `observation.images.left_wrist`: `VISUAL`, shape `[3, 480, 640]`70- `observation.images.top`: `VISUAL`, shape `[3, 480, 640]`71- `observation.state`: `STATE`, shape `[6]`72- `observation.state.radian_urdf0`: `STATE`, shape `[6]`73 74## Outputs75 76- `action`: `ACTION`, shape `[6]`77- `action.radian_urdf0`: `ACTION`, shape `[6]`78 79## Usage80 81Install and use the same LeRobot checkout/environment that contains the GR00T policy82implementation, then point `policy.path` to this Hub repo.83 84```bash85lerobot-record \86  --robot.type=<your_robot> \87  --dataset.repo_id=<your_eval_dataset_repo> \88  --policy.path=CoRL2026-CSI/IsaacLab-so101-multi-gr00t \89  --episodes=1090```91 92For local Python usage, load the policy with LeRobot's policy factory or GR00T policy93loader from the training checkout.94 95## Evaluation96 97This upload records the offline training run metrics only. No rollout success rate is98claimed here unless a separate real/sim evaluation is added later.99 100Final logged training metrics:101 102- loss: `0.010196`103- grad norm: `0.193945`104- update time: `1.2521 s/step`105- dataloading time: `0.0046 s/step`106 107## Limitations and Safety108 109This model is a robot control policy and can produce unsafe actions if deployed on110hardware without appropriate validation, workspace limits, emergency stop handling, and111task-specific safety checks. Test in simulation or a constrained setup before any112physical deployment.113 114The model is specialized to the training dataset and embodiment configuration. It may115not transfer reliably to different cameras, calibration, action spaces, robot hardware,116or tasks without further validation or fine-tuning.117 118## License and Terms119 120This model is a fine-tune of `nvidia/GR00T-N1.5-3B`; users are responsible for complying121with the NVIDIA model license and any dataset/license constraints. See the base model122card and NVIDIA license terms linked in the metadata.123 124## Files125 126- `model.safetensors`: fine-tuned policy weights127- `config.json`: LeRobot GR00T policy config128- `train_config.json`: training configuration129- `policy_preprocessor.json` and `policy_postprocessor.json`: LeRobot processor pipelines130- `policy_*_step_*.safetensors`: normalization/statistics state used by processors131