dexbench/single-lerobot

DexBench — Single-Hand

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

Summary

Episodes	735
Frames	287,019
Tasks	15
FPS	30
Camera resolution	256 × 256
Robot	Floating-base Shadow Dexterous Hand (right)
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	(28,)
observation.images.third_person	video (h264)	(256, 256, 3)
observation.images.wrist	video (h264)	(256, 256, 3)
action	float32	(28,)

Tasks

task_index	name
0	Dexbench-OpenFaucet-v0
1	Dexbench-FunctionalDrillApply-v0
2	Dexbench-FunctionalHammerStrike-v0
3	Dexbench-FunctionalPourCan-v0
4	Dexbench-FunctionalPourMug-v0
5	Dexbench-PivotLargeCuboidAgainstWall-v0
6	Dexbench-TakeBookOffShelf-v0
7	Dexbench-GraspBleach-v0
8	Dexbench-GraspCup-v0
9	Dexbench-GraspKettle-v0
10	Dexbench-GraspPan-v0
11	Dexbench-PickThinObjectFromContainer-v0
12	Dexbench-GearMesh-v0
13	Dexbench-InsertPeg-v0
14	Dexbench-PlugCharger-v0

Usage

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

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

frame = ds[0]
# frame["observation.state"], frame["action"]: torch.Tensor(28,)
# frame["observation.images.third_person"], frame["observation.images.wrist"]: torch.Tensor(3, 256, 256)
# frame["task"]: str (one of the 15 task ids above)

Companion dataset

• dexbench/bimanual-lerobot — bimanual subset (746 ep, 56-dim action, additional left/right wrist cameras).

Source

Replayed from teleop trajectory pickles in the gated dexbench/DexBench_dataset repo (teleop_dataset/{rigid, articulation, dexterous, 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.