dexbench/bimanual-lerobot

DexBench — Bimanual

Bimanual dexterous manipulation demonstrations from DexBench, replayed through Isaac Lab and packaged as a LeRobot v2.1 dataset. Each frame ships with third-person + left-wrist + right-wrist RGB video (256×256 @ 30 fps), proprioception, and the corresponding joint-space action.

Summary

Episodes	746
Frames	298,482
Tasks	9
FPS	30
Camera resolution	256 × 256
Robot	Floating-base bimanual Shadow Dexterous Hands
LeRobot codebase version	v2.1
Per-episode visual randomization	disabled (fixed HDRI background + table texture across all episodes)

Feature schema

Feature	dtype	shape
observation.state	float32	(56,)
observation.images.third_person	video (h264)	(256, 256, 3)
observation.images.left_wrist	video (h264)	(256, 256, 3)
observation.images.right_wrist	video (h264)	(256, 256, 3)
action	float32	(56,)

Tasks

task_index	name
0	Dexbench-FixateThenManipulate-LiftBasketHandle-v0
1	Dexbench-FixateThenManipulate-OpenFlatFolder-v0
2	Dexbench-FixateThenManipulate-OpenHuaweiPhone-v0
3	Dexbench-FixateThenManipulate-OpenLaptop-v0
4	Dexbench-FixateThenManipulate-OpenStapler-v0
5	Dexbench-FixateThenManipulate-SlideUtilityKnife-v0
6	Dexbench-FixateThenManipulate-SqueezeScissors-v0
7	Dexbench-BimanualLiftBasket-v0
8	Dexbench-BimanualLiftTray-v0

Some tasks have additional task-condition variants (e.g. _lose_startup, _onthetable) that share the same base task identifier; episodes from those variants are merged under the base task_index listed above.

Usage

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("dexbench/bimanual-lerobot")
print(ds.num_episodes, ds.num_frames)

frame = ds[0]
# frame["observation.state"], frame["action"]: torch.Tensor(56,)
# frame["observation.images.third_person"],
# frame["observation.images.left_wrist"],
# frame["observation.images.right_wrist"]: torch.Tensor(3, 256, 256)
# frame["task"]: str (one of the 9 task ids above)

Companion dataset

• dexbench/single-lerobot — single-hand subset (735 ep, 15 tasks, 28-dim action, one wrist camera).

Source

Replayed from teleop trajectory pickles in the gated dexbench/DexBench_dataset repo (teleop_dataset/{bimanual_articulation, grasping}/...) using DexBench's scripts/create_demo_files_sequential.py (one Isaac Sim env per task, episodes replayed serially within the env, per-episode HDRI / table-texture randomizers disabled for visually stable demos). Then converted with scripts/convert_to_lerobot.py (h264 video at CRF 18).

License

Apache-2.0. Defers to upstream DexBench terms for the underlying assets and teleop data.