DexBench (RLDS)

DexBench dexterous manipulation demonstrations replayed through Isaac Lab, packaged as Open-X-Embodiment-style RLDS / TFDS datasets. Each frame: third-person + wrist RGB at 256×256, proprioception, joint-space action, and a natural-language instruction.

This repo contains two variants as sibling subdirectories — same source data + recording pipeline, different robot embodiments → separate TFDS configs:

variant	episodes	tasks	action / state dim	wrist cameras	TFDS name
single	735	15	28	wrist_image	dexbench_rlds/single
bimanual	746	9	56	left_wrist_image + right_wrist_image	dexbench_rlds/bimanual

Total: 1,481 episodes, ~585k frames, 30 fps, 256×256 RGB.

Per-episode visual randomization is disabled — the HDRI background and table texture are fixed across all episodes for visually stable demonstrations.

Feature schema

Both variants share the same Open-X-Embodiment step structure. Image/state/action shapes vary by variant.

FeaturesDict({
    'episode_metadata': FeaturesDict({
        'file_path': Text(),
        'task': Text(),
    }),
    'steps': Dataset({
        'observation': FeaturesDict({
            'image':              Image((256, 256, 3), uint8),    # third-person
            'wrist_image':        Image((256, 256, 3), uint8),    # single only
            'left_wrist_image':   Image((256, 256, 3), uint8),    # bimanual only
            'right_wrist_image':  Image((256, 256, 3), uint8),    # bimanual only
            'state':              Tensor((28 | 56,), float32),
        }),
        'action':               Tensor((28 | 56,), float32),
        'discount':             Scalar(float32),                  # always 1.0
        'reward':               Scalar(float32),                  # 1.0 on terminal success step, else 0.0
        'is_first':             Scalar(bool),
        'is_last':              Scalar(bool),
        'is_terminal':          Scalar(bool),
        'language_instruction': Text(),
    }),
})

Usage

The repo layout matches the TFDS data_dir convention. Clone the full repo, then load with tensorflow_datasets:

mkdir -p ~/data/dexbench_rlds
hf download dexbench/rlds --repo-type dataset --local-dir ~/data/dexbench_rlds
# After: ~/data/dexbench_rlds/{single,bimanual}/1.0.0/*.tfrecord-NNNNN-of-NNNNN

import tensorflow_datasets as tfds

# Single-hand variant
ds_s = tfds.builder("dexbench_rlds/single",   data_dir="~/data").as_dataset(split="train")
# Bimanual variant
ds_b = tfds.builder("dexbench_rlds/bimanual", data_dir="~/data").as_dataset(split="train")

for ep in ds_s.take(1):
    print(ep["episode_metadata"]["task"].numpy().decode())
    for step in ep["steps"].take(1):
        img    = step["observation"]["image"]         # (256, 256, 3) uint8
        state  = step["observation"]["state"]         # (28,) float32
        action = step["action"]                       # (28,) float32
        instr  = step["language_instruction"].numpy().decode()

Tasks

Single-hand (15)

task_index	task identifier	language instruction
0	Dexbench-OpenFaucet-v0	"open the faucet"
1	Dexbench-FunctionalDrillApply-v0	"operate the power drill"
2	Dexbench-FunctionalHammerStrike-v0	"strike the nail with the hammer"
3	Dexbench-FunctionalPourCan-v0	"pour from the can"
4	Dexbench-FunctionalPourMug-v0	"pour from the mug"
5	Dexbench-PivotLargeCuboidAgainstWall-v0	"pivot the large cuboid against the wall"
6	Dexbench-TakeBookOffShelf-v0	"take the book off the shelf"
7	Dexbench-GraspBleach-v0	"grasp the bleach bottle"
8	Dexbench-GraspCup-v0	"grasp the cup"
9	Dexbench-GraspKettle-v0	"grasp the kettle"
10	Dexbench-GraspPan-v0	"grasp the pan"
11	Dexbench-PickThinObjectFromContainer-v0	"pick the thin object out of the container"
12	Dexbench-GearMesh-v0	"mesh the gears together"
13	Dexbench-InsertPeg-v0	"insert the peg into the hole"
14	Dexbench-PlugCharger-v0	"plug the charger into the receptacle"

Bimanual (9)

task_index	task identifier	language instruction
0	Dexbench-FixateThenManipulate-LiftBasketHandle-v0	"lift the basket by its handle"
1	Dexbench-FixateThenManipulate-OpenFlatFolder-v0	"open the flat folder"
2	Dexbench-FixateThenManipulate-OpenHuaweiPhone-v0	"open the phone case"
3	Dexbench-FixateThenManipulate-OpenLaptop-v0	"open the laptop"
4	Dexbench-FixateThenManipulate-OpenStapler-v0	"open the stapler"
5	Dexbench-FixateThenManipulate-SlideUtilityKnife-v0	"slide out the utility knife blade"
6	Dexbench-FixateThenManipulate-SqueezeScissors-v0	"squeeze the scissors"
7	Dexbench-BimanualLiftBasket-v0	"lift the basket with both hands"
8	Dexbench-BimanualLiftTray-v0	"lift the tray with both hands"

Some bimanual tasks have condition variants (_lose_startup, _onthetable) merged under the base task id.

LeRobot mirror

The same source data is also released as LeRobot v2.1 datasets, useful if you prefer that loader:

• dexbench/single-lerobot — single-hand
• dexbench/bimanual-lerobot — bimanual

Source

Replayed from teleop trajectory pickles in the gated dexbench/DexBench_dataset repo 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). Then converted with scripts/convert_to_rlds.py (jpeg-in-tfrecord).

License

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