Datasets:

RLWRLD
/

DexBench-datasets

license: cc-by-4.0
task_categories:
  - robotics
tags:
  - LeRobot
  - dexterous-manipulation
  - bimanual
  - mobile-manipulation
configs:
  - config_name: multitask
    data_files: multitask/data/*/*.parquet
    default: true
  - config_name: db_r1
    data_files: db_r1/data/*/*.parquet
  - config_name: db_r2
    data_files: db_r2/data/*/*.parquet
  - config_name: db_r3
    data_files: db_r3/data/*/*.parquet
  - config_name: db_l1
    data_files: db_l1/data/*/*.parquet
  - config_name: db_l2
    data_files: db_l2/data/*/*.parquet
  - config_name: db_l3
    data_files: db_l3/data/*/*.parquet
  - config_name: db_l4
    data_files: db_l4/data/*/*.parquet
  - config_name: db_m1
    data_files: db_m1/data/*/*.parquet
  - config_name: db_m2
    data_files: db_m2/data/*/*.parquet
  - config_name: db_m3
    data_files: db_m3/data/*/*.parquet
  - config_name: db_m4
    data_files: db_m4/data/*/*.parquet
  - config_name: db_m5
    data_files: db_m5/data/*/*.parquet
  - config_name: db_h1
    data_files: db_h1/data/*/*.parquet
  - config_name: db_h2
    data_files: db_h2/data/*/*.parquet
  - config_name: db_h3
    data_files: db_h3/data/*/*.parquet

DexBench datasets

Demonstrations for the 15 DexBench tasks: 240 episodes per task, 3,600 episodes and 2,947,227 frames in all. Robot: RB-Y1 mobile manipulator with two WUJI dexterous hands (66 joints).

Each task folder is one LeRobot v3.0 dataset with its own README; multitask/ holds all 15 tasks in one dataset (3,600 episodes, task_index 0–29, one instruction per episode).

Folder	Task	Name	Episodes	Frames
`db_r1/`	DB-R1	Pick fruit from a display	240	51,239
`db_r2/`	DB-R2	Retrieve a drink from a refrigerator	240	120,813
`db_r3/`	DB-R3	Place a box in a shelf slot	240	252,802
`db_l1/`	DB-L1	Pick polybags from a conveyor	240	187,312
`db_l2/`	DB-L2	Lift a package onto a conveyor	240	79,158
`db_l3/`	DB-L3	Pack an invoice	240	65,367
`db_l4/`	DB-L4	Deliver a parcel barcode-up	240	308,559
`db_m1/`	DB-M1	Insert a detergent drawer	240	354,546
`db_m2/`	DB-M2	Insert a connector	240	240,697
`db_m3/`	DB-M3	Fasten bolts with a power driver	240	247,762
`db_m4/`	DB-M4	Hand-start a leveling leg	240	403,558
`db_m5/`	DB-M5	Place a part in a kit box	240	103,610
`db_h1/`	DB-H1	Stack plates or bowls	240	187,019
`db_h2/`	DB-H2	Lift and pour	240	301,434
`db_h3/`	DB-H3	Extract one utensil	240	43,351
`multitask/`	all	all 15 tasks	3,600	2,947,227

Contents of every dataset

Cameras: observation.images.camera_ego_left, observation.images.camera_ego_right, observation.images.camera_wrist_left, observation.images.camera_wrist_right: 640×360, H.264, 30 fps.
State / action: observation.joint_position, observation.joint_velocity and action: float32 × 66, joint names in meta/info.json, groups in meta/modality.json: wheels 0:4, torso 4:10, right arm 10:17, right hand 17:37, left arm 37:44, left hand 44:64, neck 64:66. Also timestamp, frame_index, episode_index, index, task_index, next.done.

Download

hf download RLWRLD/DexBench-datasets --repo-type dataset --include "db_r1/*" "tools/*" --local-dir dexbench

Load (LeRobot v3.0; tested with lerobot 0.4.4)

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("db_r1", root="dexbench/db_r1")

Convert to LeRobot v2.1 (tested with lerobot 0.3.3 reading the result)

pip install "av>=14" numpy pandas pyarrow
cd dexbench
python tools/convert_lerobot_v30_to_v21.py db_r1 db_r1_v21 --expect-episodes 240 --expect-frames 51239

The conversion is lossless: parquet rows are copied unchanged, each episode's video is cut out of the concatenated file without re-encoding and checked frame by frame, and meta/stats.json, meta/episodes_stats.jsonl, meta/episodes.jsonl and meta/tasks.jsonl are written for v2.1 readers. Each folder's README gives the command with its episode and frame counts.

Integrity

Every dataset folder has a SHA256SUMS over its files: cd db_r1 && sha256sum -c SHA256SUMS.