andreaskoepf/robotwin_dk1_realcam_randomized_100

This dataset was created using LeRobot.

Dataset Description

Multi-task bimanual manipulation dataset combining 50 tasks (5000 episodes total), generated in the RoboTwin simulator with domain randomization. Episodes are shuffled across tasks for mixed-task training.

• Robot: TRLC DK1 bimanual (bi_dk1_follower) - 2x 6-DOF arms with parallel grippers
• Tasks: 50 distinct manipulation tasks
• Episodes: 5000 (shuffled across tasks)
• Total frames: 772,264 (757.1 min @ 17 fps)

Cameras

Camera	Resolution	Codec	FPS
observation.images.head	640x360	av1	17
observation.images.left_wrist	640x360	av1	17
observation.images.right_wrist	640x360	av1	17

Action & State Space

• action: float32[14] - joint position targets (6 joints + 1 gripper per arm)
• observation.state: float32[14] - current joint positions + gripper state

Each episode has a natural language task description generated from RoboTwin templates.

Task Distribution

Task	Episodes
Adjust Bottle	100
Beat Block Hammer	100
Blocks Ranking Rgb	100
Blocks Ranking Size	100
Click Alarmclock	100
Click Bell	100
Dump Bin Bigbin	100
Grab Roller	100
Handover Block	100
Handover Mic	100
Hanging Mug	100
Lift Pot	100
Move Can Pot	100
Move Pillbottle Pad	100
Move Playingcard Away	100
Move Stapler Pad	100
Open Laptop	100
Open Microwave	100
Pick Diverse Bottles	100
Pick Dual Bottles	100
Place A2B Left	100
Place A2B Right	100
Place Bread Basket	100
Place Bread Skillet	100
Place Burger Fries	100
Place Can Basket	100
Place Cans Plasticbox	100
Place Container Plate	100
Place Dual Shoes	100
Place Empty Cup	100
Place Fan	100
Place Mouse Pad	100
Place Object Basket	100
Place Object Scale	100
Place Object Stand	100
Place Phone Stand	100
Place Shoe	100
Press Stapler	100
Put Bottles Dustbin	100
Put Object Cabinet	100
Rotate Qrcode	100
Scan Object	100
Shake Bottle	100
Shake Bottle Horizontally	100
Stack Blocks Three	100
Stack Blocks Two	100
Stack Bowls Three	100
Stack Bowls Two	100
Stamp Seal	100
Turn Switch	100

Episode Manifest

The file manifest.json maps each dataset episode index to its source task and original episode number:

[
  {"dataset_episode": 0, "task": "shake_bottle", "original_episode": 42},
  {"dataset_episode": 1, "task": "place_bread_basket", "original_episode": 7},
  ...
]

Loading the Dataset

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("andreaskoepf/robotwin_dk1_realcam_randomized_100")
frame = dataset[0]
print(frame["observation.state"].shape)  # torch.Size([14])
print(frame["action"].shape)             # torch.Size([14])

Citation

@misc{robotwin2025,
  title = {RoboTwin: Dual-Arm Robot Benchmark with Generative Digital Twins},
  author = {Chen, Yao Mu et al.},
  year = {2025},
  url = {https://github.com/TianxingChen/RoboTwin}
}