datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
RMBench-taco-gemini
RMBench-taco-gemini
RMBench training episodes (9 tasks, 449 episodes, 30 fps) with dense high-level labels produced by the TACOR offline annotator:
Gemini 3.7 Flash reads each whole episode as one video clip (one sample every 25 frames) and labels every sampled frame under the
task-specific context (taco) induced for that task. Each tick carries the current subtask, the running textual memory and the
visual-memory operations (keyframe store / retrieval) that the online… See the full description on the dataset page: https://huggingface.co/datasets/Myungkyu/RMBench-taco-gemini.RMBench-preset-gemini
RMBench-preset-gemini
RMBench training episodes (9 tasks, 450 episodes, 30 fps) with dense high-level labels produced by the TACOR offline annotator:
Gemini 3.7 Flash reads each whole episode as one video clip (one sample every 25 frames) and labels every sampled frame given only
the task's subtask preset (the ordered list of subtask labels, no further task-specific guidance). Each tick carries the current
subtask, the running textual memory and the visual-memory operations… See the full description on the dataset page: https://huggingface.co/datasets/Myungkyu/RMBench-preset-gemini.RMBench-preset-luna
RMBench-preset-luna
RMBench training episodes (9 tasks, 450 episodes, 30 fps) with dense high-level labels produced by the TACOR offline annotator:
GPT-5.6 Luna (gpt-5.6-luna) reads the frames of each episode sampled every 25 frames as labelled images and labels every sampled frame given only
the task's subtask preset (the ordered list of subtask labels, no further task-specific guidance). Each tick carries the current
subtask, the running textual memory and the visual-memory… See the full description on the dataset page: https://huggingface.co/datasets/Myungkyu/RMBench-preset-luna.RMBench-taco-wodemo-gemini
RMBench-taco-wodemo-gemini
RMBench training episodes (9 tasks, 450 episodes, 30 fps) with dense high-level labels produced by the TACOR offline annotator:
Gemini 3.7 Flash reads each whole episode as one video clip (one sample every 25 frames) and labels every sampled frame under a
task-specific context (taco) for that task. Each tick carries the current subtask, the running textual memory and the
visual-memory operations (keyframe store / retrieval) that the online high-level… See the full description on the dataset page: https://huggingface.co/datasets/Myungkyu/RMBench-taco-wodemo-gemini.RMBench-labeled_newMem-0-m1mix-dataset-RMBench
Mem-0 m1_mix — RMBench / RoboTwin 2.0 (LeRobot dataset)
The m1_mix training dataset for the Mem-0 execution module: the five RMBench
M1 tasks merged into a single LeRobot
v2.1 dataset with globally unique episode indices. This is the exact data used to
train the checkpoint released at
qiuly/Mem-0-m1mix-RMBench.
Summary
Format
LeRobot v2.1
Episodes
250 (50 per task × 5 tasks)
Frames
92,520
FPS
30
Tasks
5 (see below)
Robot
dual-arm (2× 7-DoF +… See the full description on the dataset page: https://huggingface.co/datasets/qiuly/Mem-0-m1mix-dataset-RMBench.rmbench_lerobot
RMBench — LeRobot (GR00T-compatible), 9 official tasks
A GR00T-compatible LeRobot v2.1 conversion of RMBench, a dual-arm robotic memory manipulation benchmark. 450 episodes (9 tasks × 50 expert demos), 277,350 frames, 3 camera views, 14-D absolute-joint action.
Source
Derived from the official RMBench release TianxingChen/RMBench (demo_clean split): RoboTwin-2.0 simulator, dual-arm Aloha-AgileX (6-DoF arm + 1 gripper per arm), per-episode HDF5, 50 expert… See the full description on the dataset page: https://huggingface.co/datasets/Myungkyu/rmbench_lerobot.RMBench_labeled
RMBench Labeled
LeRobot-format demonstrations for the nine RMBench tasks, including manually
reviewed keyframe interval annotations in each task's meta/keyframes.json.
Each task is stored as <task>/{data,meta,videos}. Eight tasks contain 50
episodes; battery_try contains 49 episodes because one invalid source episode
was removed.
The MP4 container metadata is 30 FPS, while each stored dataset timestep
corresponds to the benchmark's approximately 16.67 Hz environment step.
rmbench_lerobot
SimpleMemVLA Dataset
This repository contains the training data for SimpleMemVLA, a vision-language-action model that processes timestamped video history as native context for long-horizon robotic manipulation. The datasets are provided in LeRobot v3 format and cover the benchmarks: RMBench, RoboMME, MIKASA-Robo, RoboMemArena, and LIBERO.
Paper: SimpleMemVLA: A Simple but Effective Native-Video Memory for Vision-Language-Action Models
Project page: Hugging Face collection
Code:… See the full description on the dataset page: https://huggingface.co/datasets/yinchenghust/rmbench_lerobot.RMBench-taco-luna
RMBench-taco-luna
RMBench training episodes (9 tasks, 449 episodes, 30 fps) with dense high-level labels produced by the TACOR offline annotator:
GPT-5.6 Luna (gpt-5.6-luna) reads the frames of each episode sampled every 25 frames as labelled images and labels every sampled frame under the
task-specific context (taco) induced for that task. Each tick carries the current subtask, the running textual memory and the
visual-memory operations (keyframe store / retrieval) that the… See the full description on the dataset page: https://huggingface.co/datasets/Myungkyu/RMBench-taco-luna.RMBench-taco-wodemo-luna
RMBench-taco-wodemo-luna
RMBench training episodes (9 tasks, 450 episodes, 30 fps) with dense high-level labels produced by the TACOR offline annotator:
GPT-5.6 Luna (gpt-5.6-luna) reads the frames of each episode sampled every 25 frames as labelled images and labels every sampled frame under a
task-specific context (taco) for that task. Each tick carries the current subtask, the running textual memory and the
visual-memory operations (keyframe store / retrieval) that the online… See the full description on the dataset page: https://huggingface.co/datasets/Myungkyu/RMBench-taco-wodemo-luna.rmbench-cover-blocks-smolvla-ttt-stage1plus2-videos
RMBench Cover Blocks: SmolVLA-TTT Stage1+2 videos
Four qualitative episodes from the Stage1+2 checkpoint. Each episode uses an independent model server so TTT fast state cannot leak between samples.
Videos use the official demo_clean configuration with unseen instructions. They are qualitative samples; the separate canonical 100-episode formal evaluation is recorded in metadata.json.
Video
GPU
Canonical seed
Success
Instruction
Open video 1
0
100000
no
On the table… See the full description on the dataset page: https://huggingface.co/datasets/QRP123/rmbench-cover-blocks-smolvla-ttt-stage1plus2-videos.RMBench-LeRobotlerobot-RMBench-Observe_n_Pickup-landmarklerobot-RMBench-CoverBlocks-landmarkRMBench-lerobot
RMBench-lerobot
RMBench LeRobot single-dataset layout with local artifacts.
This public repository was uploaded from local research artifacts under
/data/nichaojun/3-data. Download metadata directories such as .hfd and
credential-bearing helper files are intentionally excluded.
Excluded from this upload batch:
/data/nichaojun/3-data/VAD
/data/nichaojun/3-data/ReconDreamer-RL/assets/nus
RMBench-lerobot-by-task
RMBench-lerobot-by-task
RMBench LeRobot by-task split with local geometry/text artifacts.
This public repository was uploaded from local research artifacts under
/data/nichaojun/3-data. Download metadata directories such as .hfd and
credential-bearing helper files are intentionally excluded.
Excluded from this upload batch:
/data/nichaojun/3-data/VAD
/data/nichaojun/3-data/ReconDreamer-RL/assets/nus
Xiang_Float_After_Tomorrow_Head_SPLITED_RMBG_MASKRMBench-DataXiang_Float_After_Tomorrow_Head_RMBG_SPLITED_VACE_OutPainting_videosRM_Base_40_2Xiang_Float_After_Tomorrow_Head_SPLITED_RMBGXiang_Float_After_Tomorrow_Head_SPLITED_RMBG_mask_videosXiang_Float_After_Tomorrow_Head_SPLITED_RMBG_placed_videosXiang_Float_After_Tomorrow_Head_RMBG_SPLITED_VACE_OutPainting_masksXiang_Float_After_Tomorrow_Head_RMBG_SPLITED_VACE_OutPainting_srcs