You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

DextraVerse rig grasp, 261 demonstrations

Teleoperated grasp demonstrations recorded on the real DextraVerse rig: a Vega dual-arm robot with two 22-DOF Sharpa dexterous hands, a head camera, two wrist cameras and fingertip tactile sensing. This is the merged corpus rig_grasp_merged_20260908_n261.

Demonstrations 261
Frames 138,382
Duration 123.0 min at 20 Hz (control loop 50 Hz)
Size 42.9 GB, one gzip-compressed HDF5 file
Hand used left 140, right 116, both 3, none 2
Recorded 2026-09-02 to 2026-09-06, ten sessions

Files

File Contents
rig_grasp_merged_20260908_n261.hdf5 The data.
rig_grasp_merged_20260908_n261.md Generated report: source sessions, observation keys, one row per demo with flags.
rig_grasp_merged_20260908_n261.json The same report, machine readable.

SHA-256 of the HDF5 file: 9961d734cad21391679c9bd5f0622fc4fd37f0c0320402fe00c4b973409d4e17

Layout

robomimic-style HDF5. Demos are data/demo_0 to data/demo_260; each has a num_samples attribute. data.attrs["env_args"] is a JSON string with the recording settings and joint names.

Dataset Per-step shape dtype
actions (63,) float32
rewards () float32
dones () int64
obs/arm_joint_pos (14,) float32
obs/arm_joint_vel (14,) float32
obs/cfg_joint_pos (63,) float32
obs/ee_pose_quat (2, 7) float32
obs/hand_joint_pos (2, 22) float32
obs/hand_joint_cmd (2, 22) float32
obs/finger_force_mag (10,) float32
obs/finger_force_dir_sensor (10, 3) float32
obs/tactile_map (10, 24, 24) uint8
obs/tacmap_sensor_pose (10, 4, 4) float32
obs/grasp_phase (2,) float32
obs/head_camera_rgb (240, 320, 3) uint8
obs/left_wrist_rgb (240, 320, 3) uint8
obs/right_wrist_rgb (240, 320, 3) uint8
obs/timestamp (1,) float32

Arrays with a leading 2 are ordered left, right. Arrays with a leading 10 cover the ten fingers. Forces are in newtons.

import h5py

with h5py.File("rig_grasp_merged_20260908_n261.hdf5", "r") as f:
    demo = f["data/demo_0"]
    actions = demo["actions"][:]               # (T, 63)
    head = demo["obs/head_camera_rgb"][:]      # (T, 240, 320, 3)
    force = demo["obs/finger_force_mag"][:]    # (T, 10)

Known flags

Four demonstrations carry quality flags. They were reviewed on 2026-09-08 and kept on purpose; filter them out if your use needs clean episodes.

Demo Flag
demo_32 4.83 s recording stall mid-episode
demo_33 5.88 s recording stall, and no grasp
demo_42 1.16 s recording stall
demo_96 no grasp

The file has no held-out split. The 49 demos at indices 212 to 260 come from one session (session_20260904_150927) and can serve as a test set against the first 212.

Downloads last month
27