1duahcua9/GRASP-OakInk2

GRASP OakInk2 Retargeting Solutions

Physics-based retargeting solutions for the OakInk2 bimanual manipulation dataset, produced with GRASP (Gradient-Guided Sampling for Contact-Rich Physics-Based Motion Retargeting).

Each of the 1352 OakInk2 demonstrations was attempted on three robot hands. A solution contains the states executed in simulation, the controls that produced them, the kinematic reference it was retargeted from, the MuJoCo scene, and a video.

Hand	Solutions	Successful	Success (of 1352)	Median position error (m)	Median rotation error (rad)	Videos
Sharpa Wave	1327	1272	94.1%	0.022	0.090	1327
XHand	1327	1281	94.7%	0.021	0.088	1327
Wuji Hand 2	1240	1179	87.2%	0.020	0.085	1240

A task is successful when the mean object position error is below 0.1 m and the mean object rotation error is below 0.5 rad. Errors are medians over all stored solutions, including failures.

Status. The Wuji Hand 2 batch was still running at this snapshot: 72 tasks are listed with state running and have no solution yet, so its counts are preliminary.

Quick start

import numpy as np
import pandas as pd
from huggingface_hub import snapshot_download

index = pd.read_parquet("hf://datasets/1duahcua9/GRASP-OakInk2/index/tasks.parquet")
row = index[(index.hand == "sharpa") & index.success].iloc[0]

root = snapshot_download(
    "1duahcua9/GRASP-OakInk2",
    repo_type="dataset",
    allow_patterns=[
        f"{row.path}/*",
        "assets/robots/sharpa_bimanual/*",
        *[f"assets/objects/{o}/*" for o in row.objects],
        "tools/*",
    ],
)
solution = np.load(f"{root}/{row.path}/trajectory.npz")
print(solution["executed_qpos"].shape)

To view it in MuJoCo, run python tools/replay.py <root>/<path> from the downloaded folder. Add --reference to play the kinematic reference instead.

XHand scenes need one extra step. The XHand meshes are not hosted here. Run python tools/fetch_xhand_assets.py to download them from the public SPIDER example dataset.

Layout

index/tasks.parquet              one row per hand and task, including tasks without a solution
index/summary.json               per-hand totals
data/<hand>/<task>/
    trajectory.npz               the executed solution
    kinematic.npz                the kinematic reference
    scene.xml                    MuJoCo scene with paths relative to this file
    task_info.json               scene metadata
    result.json                  solver settings and diagnostics
videos/<hand>/<task>.mp4         render of executed_qpos at 25 fps
assets/objects/<object>/         visual and convex collision meshes
assets/robots/                   Sharpa and Wuji meshes
tools/                           replay and asset download scripts

Object folders keep the OakInk2 object ID. Each task has its own copy of its object meshes, so folders ending in __v2, __v3 and so on are other copies of the same object. The objects column of the index lists the folders a task's scene needs.

Files

trajectory.npz, sampled every 5 ms (200 Hz):

Key	Shape	Meaning
executed_qpos	(T, nq)	Joint positions after each simulation step of the closed-loop run
executed_qvel	(T, nv)	Joint velocities after each simulation step
committed_steps	(T, nu)	Per-step delta added to the reference position target
committed_knots	(K, nu)	Control knots committed by the planner
committed_cost_world	(N,)	Cost diagnostics per parallel world
j_full_slots	(N,)	Cost diagnostics per parallel world
daid_pos_err_frames	(T,)	Object position error per frame (m)
daid_quat_err_frames	(T,)	Object rotation error per frame (rad)
control_interface	scalar	direct_pd_delta

kinematic.npz holds the kinematic reference from inverse kinematics: qpos, qvel and frequency (50 Hz). Frame i >= 1 of the reference aligns with executed_qpos[4 * i - 1].

The state vector is ordered as right hand, left hand, right object, left object. Each hand has six wrist degrees of freedom (three translations and three rotations) followed by its finger joints. Each object is a free joint of seven values, position then wxyz quaternion. The joint names are in each scene.xml.

Hand	nq	nv	nu	Finger joints per hand
Sharpa Wave	70	68	56	22
XHand	50	48	36	12
Wuji Hand 2	66	64	52	20

The columns of index/tasks.parquet include the task, its action name, state, success, the error metrics, the trajectory length, the object folders, and the path and video of the solution. state is completed for tasks with a solution; other values mean no solution exists. excluded_from_stats flags tasks with an invalid scene or an outlier result, and exclusion_reason gives the reason. The table above still counts these tasks.

Limitations

• Use executed_qpos as the result. The solutions were produced by closed-loop planning in MuJoCo Warp, which is not deterministic on contact-rich tasks. Replaying committed_steps open loop is not expected to reproduce them.
• The solver simulated at 5 ms. scene.xml keeps a default timestep of 2 ms, so set model.opt.timestep = 0.005 if you simulate with these controls.
• Failed solutions are included. Filter on success. A few failed runs have very large errors because an object was launched, so use medians or filter before averaging.

Licenses

Everything here is released under CC BY-SA 4.0 unless listed below. The solutions, kinematic references, videos and object meshes derive from OakInk2, published under that license on its dataset page.

Content	Terms
assets/robots/sharpa_bimanual and Sharpa robot parameters in scenes	Apache-2.0, Copyright 2025 Sharpa Group. See the included LICENSE.txt and NOTICE.txt. From sharpa-urdf-usd-xml. The visual meshes left_ and right_hand_C_MC_visual.STL, left_ and right_hand_C_MC_visual_.STL, and left_ and right_pinky_MC_visual.STL differ from the current upstream files.
assets/robots/wuji_hand2 and Wuji robot parameters in scenes	MIT, Copyright 2025 Wuji Technology. See the included LICENSE. From wuji-description.
XHand robot parameters in data/xhand/*/scene.xml	Taken from the SPIDER repository, which is CC BY-NC 4.0. XHand meshes are not redistributed here.

This dataset contains no MANO model data and no human hand meshes.

Citation

@misc{liu2026grasp,
  title  = {GRASP: Gradient-Guided Sampling for Contact-Rich Physics-Based Motion Retargeting},
  author = {Liu, Yang and Wang, Tong and Bao, Yiming and Yang, Heng},
  year   = {2026}
}

Please also cite OakInk2.