Weave: Learning Whole-Body Dexterous Loco-Manipulation from Human-Object Interactions

Reference motion and physical rollouts for Weave (Hugging Face page).

• Paper: https://arxiv.org/abs/2609.16683
• Project page: https://xiaohu-art.github.io/Weave/
• Code: https://github.com/xiaohu-art/Weave

Contents

Path	What it is	Clips	Duration
reference/{train,test}/	Interaction-preserving reference motions retargeted to the robot	9,474	23.2 h
rollout/{train,test}/	Trajectories the trained policy physically executed in simulation	8,699	21.2 h
objects/	Interaction-object assets: mesh, USD, voxel SDF, surface points, shared BPS basis	9 objects	—

Robot: Unitree G1 with 29 actuated body DoFs and two Inspire hands with 12 actuated finger DoFs. Everything is recorded at 50 fps.

Data format

Every file under reference/ and rollout/ is a compressed NumPy archive with the same 15 keys. Clips are concatenated along a single flat frame axis. T is the total frame count and N the number of clips.

Key	Shape	dtype	Meaning
joint_pos	(T, 53)	float32	joint positions, rad
joint_vel	(T, 53)	float32	joint velocities, rad/s
body_pos_w	(T, 54, 3)	float32	body positions, m
body_quat_w	(T, 54, 4)	float32	body orientations, wxyz
body_lin_vel_w	(T, 54, 3)	float32	body linear velocities, m/s
body_ang_vel_w	(T, 54, 3)	float32	body angular velocities, rad/s
object_pos_w	(T, 3)	float32	object position, m
object_quat_w	(T, 4)	float32	object orientation, wxyz
object_lin_vel_w	(T, 3)	float32	object linear velocity, m/s
object_ang_vel_w	(T, 3)	float32	object angular velocity, rad/s
contact_label	(T, 54)	float32	per-body contact annotation, see below
motion_lengths	(N,)	int64	frame count of each clip
motion_names	(N,)	unicode	clip identifiers
object_names	(N,)	unicode	object each clip interacts with
fps	(1,)	int64	50 for every file

contact_label differs between the two sets

In reference/ the label is three-valued and carries both positive and negative supervision:

Value	Meaning
1	the body should be in contact with the object
-1	the body should not be in contact
0	unconstrained

In rollout/ the label is the measured contact state of the executed trajectory, thresholded at 1 N, so it takes only 0 and 1. Code that consumes the reference labels with a label != 0 mask will therefore treat every non-contact frame of a rollout as unconstrained. Convert before mixing the two sets.

Joint and body ordering

The 53 joints and 54 bodies follow the Isaac Lab articulation order of the G1 + Inspire URDF. Index 0 of the body axis is pelvis.

Body order (54)

pelvis, left_hip_pitch_link, right_hip_pitch_link, waist_yaw_link,
left_hip_roll_link, right_hip_roll_link, waist_roll_link, left_hip_yaw_link,
right_hip_yaw_link, torso_link, left_knee_link, right_knee_link,
left_shoulder_pitch_link, right_shoulder_pitch_link, left_ankle_pitch_link,
right_ankle_pitch_link, left_shoulder_roll_link, right_shoulder_roll_link,
left_ankle_roll_link, right_ankle_roll_link, left_shoulder_yaw_link,
right_shoulder_yaw_link, left_elbow_link, right_elbow_link,
left_wrist_roll_link, right_wrist_roll_link, left_wrist_pitch_link,
right_wrist_pitch_link, left_wrist_yaw_link, right_wrist_yaw_link,
L_index_proximal, L_middle_proximal, L_pinky_proximal, L_ring_proximal,
L_thumb_proximal_base, R_index_proximal, R_middle_proximal, R_pinky_proximal,
R_ring_proximal, R_thumb_proximal_base, L_index_intermediate,
L_middle_intermediate, L_pinky_intermediate, L_ring_intermediate,
L_thumb_proximal, R_index_intermediate, R_middle_intermediate,
R_pinky_intermediate, R_ring_intermediate, R_thumb_proximal,
L_thumb_intermediate, R_thumb_intermediate, L_thumb_distal, R_thumb_distal

Joint order (53)

left_hip_pitch_joint, right_hip_pitch_joint, waist_yaw_joint,
left_hip_roll_joint, right_hip_roll_joint, waist_roll_joint,
left_hip_yaw_joint, right_hip_yaw_joint, waist_pitch_joint, left_knee_joint,
right_knee_joint, left_shoulder_pitch_joint, right_shoulder_pitch_joint,
left_ankle_pitch_joint, right_ankle_pitch_joint, left_shoulder_roll_joint,
right_shoulder_roll_joint, left_ankle_roll_joint, right_ankle_roll_joint,
left_shoulder_yaw_joint, right_shoulder_yaw_joint, left_elbow_joint,
right_elbow_joint, left_wrist_roll_joint, right_wrist_roll_joint,
left_wrist_pitch_joint, right_wrist_pitch_joint, left_wrist_yaw_joint,
right_wrist_yaw_joint, L_index_proximal_joint, L_middle_proximal_joint,
L_pinky_proximal_joint, L_ring_proximal_joint, L_thumb_proximal_yaw_joint,
R_index_proximal_joint, R_middle_proximal_joint, R_pinky_proximal_joint,
R_ring_proximal_joint, R_thumb_proximal_yaw_joint, L_index_intermediate_joint,
L_middle_intermediate_joint, L_pinky_intermediate_joint,
L_ring_intermediate_joint, L_thumb_proximal_pitch_joint,
R_index_intermediate_joint, R_middle_intermediate_joint,
R_pinky_intermediate_joint, R_ring_intermediate_joint,
R_thumb_proximal_pitch_joint, L_thumb_intermediate_joint,
R_thumb_intermediate_joint, L_thumb_distal_joint, R_thumb_distal_joint

Object assets

objects/<name>/ holds the interaction object in several forms:

File	What it is
<name>.obj	source triangle mesh
<name>.usd	Isaac Sim asset, convex-decomposition collider
sdf_128.npz	signed distance field on a 128³ grid
surface.npy	surface sample points used for hand-object distance terms

clothesstand/ additionally carries a .usda, the text form of the same USD.

objects/geometry/bps_128.npy is the shared basis point set behind the BPS-SDF geometry descriptor; it is common to all objects.

Loading

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "appolyn/Weave",
    "rollout/test/woodchair_rollout.npz",
    repo_type="dataset",
)
data = np.load(path, allow_pickle=True)

lengths = data["motion_lengths"]
starts = np.concatenate([[0], np.cumsum(lengths)])

i = 0  # first clip
sl = slice(starts[i], starts[i + 1])
print(data["motion_names"][i], data["object_names"][i], lengths[i], "frames")
print(data["joint_pos"][sl].shape)      # (L, 53)
print(data["body_pos_w"][sl].shape)     # (L, 54, 3)
print(data["object_pos_w"][sl].shape)   # (L, 3)

The files are 140 MB to 1.3 GB each. np.load on an .npz is lazy, so reading motion_lengths or a single field does not decompress the rest.

Citation

@misc{cao2026weave,
  title = {{Weave}: Learning Whole-Body Dexterous Loco-Manipulation
           from Human-Object Interactions},
  author = {Liu Cao and Xingze Wu and Jingzhi Cui and Botian Xu
            and Mingzhi Pei and Ruoqu Chen and Mengdi Xu},
  year = {2026},
  eprint = {2609.16683},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2609.16683}
}