OpenMOSS-Team/robotwin2.0-lerobot-v3.0

RoboTwin 2.0 LeRobot v3.0

This dataset packages RoboTwin 2.0 bimanual-manipulation demonstrations in LeRobot v3.0 format for EasyWAM training. Clean and randomized demonstrations are stored as two independent datasets so they can be trained separately or combined. Both contain synchronized high-view and dual-wrist RGB video, bimanual joint state, actions, and natural-language task metadata.

Dataset Summary

Split	Episodes	Frames	Videos	Recording rate
Clean	2,500	549,787	7,500	50 FPS
Randomized	25,000	5,525,316	75,000	50 FPS
Full	27,500	6,075,103	82,500	50 FPS

The clean split contains 50 trajectories for each of the 50 RoboTwin tasks. The randomized split contains 500 trajectories per task. Full training uses both splits at their original 1:10 ratio.

Structure

robotwin2.0-lerobot-v3.0/
├── clean/
│   ├── data/                  # frame-level Parquet files
│   ├── meta/
│   │   ├── episodes/          # episode metadata
│   │   ├── info.json          # clean schema and totals
│   │   ├── stats.json         # clean LeRobot feature statistics
│   │   └── tasks.parquet      # clean task index and language
│   ├── videos/                # clean camera MP4 files
│   └── dataset_stats.json     # clean EasyWAM normalization statistics
├── random/
│   ├── data/
│   ├── meta/
│   │   ├── episodes/
│   │   ├── info.json
│   │   ├── stats.json
│   │   └── tasks.parquet
│   ├── videos/
│   └── dataset_stats.json     # randomized normalization statistics
└── dataset_stats.json         # full-data normalization statistics

clean/ and random/ are self-contained LeRobot v3.0 datasets with independent, zero-based episode, frame, and task indices. The root-level dataset_stats.json is used only when both datasets are trained together.

Features

Feature	Type / shape	Description
observation.images.cam_high	RGB video, 480×640	High camera
observation.images.cam_left_wrist	RGB video, 480×640	Left wrist camera
observation.images.cam_right_wrist	RGB video, 480×640	Right wrist camera
observation.state	float32[14]	Left and right arm joints and grippers
action	float32[14]	Left and right arm joint and gripper commands
timestamp	float32	Frame timestamp
frame_index	int64	Frame index within the episode
episode_index	int64	Episode identifier
task_index	int64	Index into meta/tasks.parquet

Videos use AV1 with YUV 4:2:0 pixel format and contain no audio.

Download

hf download OpenMOSS-Team/robotwin2.0-lerobot-v3.0 \
  --repo-type dataset \
  --local-dir data/robotwin2.0-lerobot-v3.0

Use with EasyWAM

The EasyWAM pipeline resizes each camera to 240×320, composes the three views into a 384×320 observation, retains a 33-step action/state horizon, and decodes 9 video timestamps. A deterministic 1% episode split is used for validation.

Select the data with data.mode=clean, data.mode=random, or data.mode=full. The default is full. The same mode applies to text-embedding precomputation, training, validation, and normalization statistics.

# Full: clean + randomized data (default)
python scripts/precompute_text_embeds.py \
  task=robotwin_easywam_mot_wan22 data.mode=full
NPROC_PER_NODE=8 bash scripts/train_zero2.sh \
  task=robotwin_easywam_mot_wan22 data.mode=full

# Clean only
python scripts/precompute_text_embeds.py \
  task=robotwin_easywam_mot_wan22 data.mode=clean
NPROC_PER_NODE=8 bash scripts/train_zero2.sh \
  task=robotwin_easywam_mot_wan22 data.mode=clean

# Randomized only
python scripts/precompute_text_embeds.py \
  task=robotwin_easywam_mot_wan22 data.mode=random
NPROC_PER_NODE=8 bash scripts/train_zero2.sh \
  task=robotwin_easywam_mot_wan22 data.mode=random

See the EasyWAM RoboTwin data guide for configuration details.

Project

• Project page: http://openmoss.ai/EasyWAM/
• Code: https://github.com/OpenMOSS/EasyWAM
• Upstream benchmark: https://github.com/robotwin-Platform/RoboTwin

License and Citation

The upstream RoboTwin 2.0 release is distributed under the MIT License. Please retain the original attribution and cite RoboTwin 2.0:

@article{chen2025robotwin,
  title   = {RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation},
  author  = {Chen, Tianxing and Chen, Zanxin and Chen, Baijun and Cai, Zijian and Liu, Yibin and Li, Zixuan and Liang, Qiwei and Lin, Xianliang and Ge, Yiheng and Gu, Zhenyu and others},
  journal = {arXiv preprint arXiv:2506.18088},
  year    = {2025}
}