INRIA's OTRetarget Uses Optimal Transport to Retarget Human Motion to Humanoids, Hitting 87% Jaccard
INRIA · hf · 2026-10-02
INRIA introduces OTRetarget, a unified method that jointly retargets human demonstration motion to both a humanoid robot and multiple objects. It models surface interactions via signed distances, closest surface points, and relative directions, then uses entropic optimal transport to transfer these quantities across human, robot, and object geometries, feeding them into a constrained inverse kinematics formulation that jointly optimizes robot and object poses per frame—without rescaling the scene. On OMOMO it achieves an 87% robot-object interaction Jaccard score and 8.7mm depth error, versus 28% and 29.3mm for OmniRetarget. The retargeted references were used to train RL whole-body policies demonstrated on a physical G1 humanoid, including two-handed box pick-and-place onto a table.
More from Embodied
- Reka Open-Sources RIDM, Extracting Camera and Motor Commands from Raw Video — RekaAILabs · 2026-10-02
- Bot Company and Physical Intelligence Cofounders Talk Robotics at Runway AI Summit — runwayml · 2026-10-02
- Local LLM User Weighs 4x MI50 vs Second 7900XTX for Bigger Models — Thac0-is-life · 2026-10-02
- Robotics-language researcher Jesse Thomason moves GLAMOR lab from USC to Georgia Tech — _jessethomason_ · 2026-10-02
- Encord's Physical AI Summit lines up Thrun, DeepMind robotics VP, Skild, Wayve — lukas_m_ziegler · 2026-10-02
- Dad uses Claude to design a 3D-printable dog toothbrush holder, full Python/SDF pipeline revealed — mbaturin · 2026-10-02