Deformable Bimanual Manipulation Benchmark
Collection
Bimanual cloth and bag manipulation in simulation: LeRobot v3.0 datasets and the policies trained on them. • 7 items • Updated
Search is not available for this dataset
video video |
|---|
grasp_silk_bimanual — bimanual silk grasping
A dual-arm Piper grasps a square of silk with both arms and lifts it. Success requires a sustained two-arm hold — a single-arm grasp counts as a failure.
This is version v4 of this dataset. Earlier iterations exist but are superseded; pin a git tag or commit sha if you need a specific version.
Part of a bimanual deformable-object (cloth / bag) manipulation benchmark. Physics is a GPU cloth solver co-simulated with the robot in a single model; observations are rendered with a photorealistic renderer.
Contents
| Format | LeRobot v3.0 (parquet + mp4, compacted) |
| Robot | sim_dual_piper_deformable |
| Episodes | 200 |
| Frames | 36,522 |
| Tasks | 1 |
| Control rate | 25 fps |
| Cameras | static_cam, right_hand_cam, left_hand_cam |
| Video | 1280×720, H264 |
| State / action | 14-D each (left 6 joints + gripper, right 6 joints + gripper) |
Notes. The cloth's initial position is randomized per episode, which makes this substantially harder than a fixed-position variant and separates policies that generalize from ones that memorize a trajectory.
Usage
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("hwk0809/grasp-silk-bimanual")
If your torchcodec build cannot decode these videos, pass video_backend="pyav".
License
Apache-2.0.
- Downloads last month
- 44