DexWeave: Learning Dexterous Humanoid Loco-Manipulation from Human Demonstrations

Paper · Project · GitHub

DexWeave studies how to transfer coordinated human body, hand, and object interactions to a dexterous humanoid. It combines interaction-consistent motion retargeting with anatomy-aware whole-body policy learning.

This dataset provides 2,070 retargeted reference sequences derived from HUMOTO and GRAB for Unitree G1 with Inspire RH56DFX hands, together with matching per-frame evaluation data. The references contain robot motion, object trajectories, and interaction annotations.

Dataset

Source	Sequences	Frames	Motion NPZ	Evaluation NPZ
HUMOTO	735	390,988	735	735
GRAB	1,335	676,647	1,335	1,335
Total	2,070	1,067,635	2,070	2,070

All sequences are exported at 50 Hz (dt = 0.02 s). The 4,140 NPZ files occupy approximately 12.9 GB.

The target robot has 29 independent body joints and 12 independent hand joints. The saved trajectories contain 53 joint coordinates, including 12 coupled joints. Use joint_names to interpret the columns of joint_pos.

These are retargeted motion references rather than policy execution rollouts. The release does not define train, validation, or test splits.

File structure

Each sequence has two matching files:

• <sequence>_opt.npz: the final robot and scene trajectory, with contact and support information.
• <sequence>_val.npz: per-frame evaluation data associated with that trajectory, including source-target hand errors and interaction masks. val is not a validation-set split.

data/
├── HUMOTO/
│   └── <sequence>/
│       ├── <sequence>_opt.npz
│       └── <sequence>_val.npz
└── GRAB/
    └── <subject>/
        └── <sequence>/
            ├── <sequence>_opt.npz
            └── <sequence>_val.npz

For GRAB, retain the subject in the sequence identifier, for example s9/wineglass_pass_1.

NPZ format

For motion files, T denotes the sequence length and N the number of scene objects.

Key	Shape	Description
root_pos	(T, 3)	Robot floating-base position in world coordinates
root_quat	(T, 4)	Robot floating-base orientation, wxyz
joint_pos	(T, 53)	Joint positions in the order specified by joint_names
joint_names	(53,)	Joint names corresponding to the saved columns
object_names	(N,)	Scene-object names
object_pos	(T, N, 3)	Object positions in world coordinates
object_quat	(T, N, 4)	Object orientations, wxyz
object_demo_scale	(T, N)	Per-frame object scales
object_active	(T, N)	Object activity masks
object_mesh_paths	(N,)	Object mesh references
contact_truth_left, contact_truth_right	(T, N, 5)	Per-object, per-finger contact annotations
fps, dt	Scalars	Export frame rate and time step
retime_source_indices	(T,)	Correspondence to the source motion timeline
source_dataset, source_seq_name	String scalars	Source dataset and sequence identifier

Positions are in meters, joint angles in radians, and quaternions use wxyz order. Robot and object poses share the saved world coordinate frame.

Evaluation files retain the matching sequence identity and time alignment. Their per-frame hand diagnostics include primary-finger position error, secondary-finger position error, palm-normal angular error, and left/right interaction masks.

Additional fields vary across sequences and source datasets; inspect data.files for the complete key list. Mesh references do not embed geometry: robot models and object meshes are not included in this NPZ release.

Download

Download all motion and evaluation files:

python -m pip install -U huggingface_hub numpy
hf download naichuan04/DexWeave --repo-type dataset --local-dir DexWeave

To download one sequence together with its matching evaluation file:

hf download naichuan04/DexWeave --repo-type dataset --local-dir DexWeave \
  --include "data/GRAB/s9/wineglass_pass_1/*"

The download preserves the directory structure. Both motion and evaluation NPZ files are provided here; reading them does not require downloading the original HUMOTO or GRAB human-motion data.

Load a sequence

from pathlib import Path
import numpy as np

folder = Path("DexWeave/data/GRAB/s9/wineglass_pass_1")

with np.load(folder / "wineglass_pass_1_opt.npz", allow_pickle=False) as motion:
    print(motion["source_seq_name"].item())       # s9/wineglass_pass_1
    print(motion["joint_pos"].shape)             # (333, 53)
    print(motion["object_pos"].shape)            # (333, 2, 3)
    print(float(motion["fps"]))                 # 50.0

with np.load(folder / "wineglass_pass_1_val.npz", allow_pickle=False) as evaluation:
    print(evaluation.files)

Keep each opt/val pair together when using the saved evaluation data.

Citation

If you use these references, please cite DexWeave:

@article{sun2026dexweave,
  author  = {Sun, Naichuan and Shen, Haotian and Zhang, Yizhang and Feng, Luying and Wang, Haoze and Xiangli, Yuanbo and Jin, Yaochu and Liu, Peidong},
  title   = {DexWeave: Learning Dexterous Humanoid Loco-Manipulation from Human Demonstrations},
  journal = {arXiv},
  year    = {2026},
}

Please also acknowledge the source datasets used in your experiments: HUMOTO, GRAB, and ContactDB, which provides object meshes used by GRAB.