AIAny
Icon for item

HOI-Retarget — Humanoid Human–Object Interaction Motion

Generates humanoid robot motion references that preserve object contact locations/timing by solving windowed trajectory optimizations against contact targets in the object frame. Releases retargeted trajectories for two Unitree robots across 75 objects (≈13.9k robot–motion pairs); CC BY‑NC‑SA 4.0.

Introduction

Why contact matters

Robotic retargeting that only matches poses often drops or penetrates the object once physics are enabled — contact geometry and timing are the missing link. HOI‑Retarget makes contact the primary constraint: every labeled contact is expressed in the object's frame and used as a target in a windowed trajectory optimizer so the resulting humanoid trajectories actually recover the intended grasps and supports under the robot's kinematic limits.

What Sets It Apart
  • Contact‑centric retargeting: contacts live in object coordinates and are treated as explicit targets during optimization, enabling the same demonstration to re-solve correctly when the object is resized or when different robot embodiments are used.
  • Large curated corpus: 6,952 distinct motions retargeted to two robots (Unitree G1 and H2) producing 13,904 robot–motion rows, covering 75 distinct object meshes and ≈13.8 hours of motion data; most rows include per-link contact flags and rendered preview videos.
  • Practical solver design: a windowed trajectory optimization balances contact satisfaction, body tracking, foot support and smoothness under joint limits (and outputs QC flags that mark wrist pinning, excessive trunk folding, or other failure modes).
  • Multi-source and extendable: assembled from five HOI capture datasets (OMOMO, ParaHome, NeuralDome, CoRoleHOI, IMHD²), supports collaborative two‑robot clips, and the pipeline can accept contact reconstructions from monocular video or new object meshes.
  • Reproducible assets and licensing: code and robot models are BSD-3-Clause; trajectories and derived assets are released under CC BY-NC-SA 4.0 (non-commercial, share-alike), with OMOMO object meshes redistributed under MIT.
Who it's for — and trade-offs

Great fit if you need realistic, contact-aware humanoid interaction references for embodied policy training, imitation learning, or simulation-based validation of whole‑body manipulation and loco‑manipulation. The dataset is especially useful when contact timing/locations must survive scaling across object sizes or robot morphologies.

Look elsewhere if you require permissive commercial licensing (the corpus is CC BY‑NC‑SA), need fine-finger grasp annotations (the robots use palm/contact links rather than per‑finger grasp detail), or require raw human mocap only (these are retargeted, robot‑centric trajectories and thus derivative of source captures).

Practical notes
  • Columns include per-frame joint angles, floating base pose, object 6‑DoF poses, per‑link contact flags, and a small preview video per row; the repo provides schema docs and a Hugging Face 3D viewer for clip playback.
  • QC metrics are provided per (motion, robot) pair; failing rows are retained but flagged so users can filter for the curated subset.
  • Because contacts are encoded in the object frame, the same clip can be re‑solved at different object scales or retargeted to additional robots with the same pipeline.

Information

  • Websitehuggingface.co
  • OrganizationsRobotic Systems Lab, ETH Zürich
  • AuthorsJihwan Shin, Adrià López Escoriza, Junzhe He, Matthias Heyrman, Marco Hutter
  • Published date2026/09/18

More Items

AI Others2026

Diagnoses whether robotic manipulation failures stem from history-based reasoning, precise physical execution, or their closed-loop coordination. It spans 30 tasks, three difficulty levels, and two scene settings.

Hugging Face

Tracks autonomous research agents publishing studies, discussing ideas, and reviewing one another over twelve weeks. Includes papers, versions, messages, and artifacts; the agents’ reported findings have not been independently verified.

Scales verifiable gameplay data by having a Builder and visual and coding players co-develop web games. Progress-weighted training improves agent performance on both its own benchmark and an external game benchmark.