ETH's HOI-Retarget cuts humanoid contact error from 18.3cm to 0.5cm, releases 6,952 open motion clips
ChongZzZhang · x · 2026-09-30
ETH Zürich's Robotic Systems Lab (Marco Hutter's group) presents HOI-Retarget, a contact-centric retargeting method that transfers human-object interaction demonstrations onto humanoid robots for large-scale motion-data generation.
Method:
- Takes mocap or monocular-video-reconstructed clips with object trajectories and contact labels
- IK retargeting maps the human to the robot; object meshes scale by height ratio, carrying contact targets
- Windowed trajectory optimization recovers contacts under the robot's kinematic limits while balancing body tracking, foot support, and smoothness
Results:
- Mean contact-point gap of 0.5cm, down from 18.3cm for the closest baseline
- 4.6× faster retargeting than the interaction-aware baseline
- Generalizes across object sizes, absorbs video-reconstructed contacts, supports multiple robots manipulating one object
Code and a dataset of 6,952 robot HOI clips (13.8 hours, 75 unique objects) are publicly released.
More from Embodied
- OpenRoboto Shift sells real-work data to fund robot training on Bittensor SN80 — markjeffrey · 2026-09-30
- OpenAI's Dots device rumored to require a Pro 200 subscription plan — testingcatalog · 2026-09-30
- Linus Ekenstam 3D prints a full skateboard quarter pipe at home — LinusEkenstam · 2026-09-30
- Qualcomm's Kedar Kondap on X2 Elite: new Surface devices, Linux support and the PC landscape — ryanshrout · 2026-09-30
- Dyna Robotics unveils Dyna-2.1, a semi-humanoid agent that completes an hour-long laundry workflow uncut — JasonMa2020 · 2026-09-30
- Single distilled policy drives ANYmal D to 3.0 m/s on flat and 2.5 m/s on rough terrain — breadli428 · 2026-09-30