Skip to content

Dataset · flex-pi

3D dual-arm robotics data from RoboTwin 2.0

flex-pi/robotwin_3d

6.18 million RGB+Depth frames and over a million text instructions for training dual-arm robotic manipulators.

What's inside

The dataset contains 27,500 episodes (6,183,813 frames) across 50 manipulation tasks from the RoboTwin 2.0 simulator for a dual-arm ALOHA-type robot (14 degrees of freedom, 50 Hz control frequency). Records are saved in the LeRobot v2.1 format and include video from 3 cameras (240x320) with depth maps (RGB + Depth) and 1,039,891 unique text instructions.

What you can do with it

  • Training Vision-Language-Action (VLA) models to control dual-arm manipulators via text commands
  • Training robotic grasping and manipulation policies using depth maps
  • Benchmarking imitation learning algorithms in robotics simulators
Tasks
Robotics
Size
1M–10M rows
Downloads this month
65.4K
Likes
1

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("flex-pi/robotwin_3d")