OTRetarget: Joint Robot and Object Motion Retargeting via Optimal Transport
The paper retargets both humanoid motion and object motion together, instead of keeping the scene fixed.
OTRetarget uses surface distances, closest points, and relative directions to preserve contacts across different human, robot, and object geometries. It folds those interaction targets into constrained inverse kinematics, jointly optimizing robot and object poses frame by frame. On OMOMO, the authors report an 87% robot-object interaction Jaccard score and 8.7 mm depth error, versus 28% and 29.3 mm for OmniRetarget. They also show transfer to a physical G1 humanoid using reinforcement-learned whole-body policies, including a two-handed box pick-and-place task. HF Daily Papers' note
OTRetarget uses surface distances, closest points, and relative directions to preserve contacts across different human, robot, and object geometries. It folds those interaction targets into constrained inverse kinematics, jointly optimizing robot and object poses frame by frame. On OMOMO, the authors report an 87% robot-object interaction Jaccard score and 8.7 mm depth error, versus 28% and 29.3 mm for OmniRetarget. They also show transfer to a physical G1 humanoid using reinforcement-learned whole-body policies, including a two-handed box pick-and-place task. HF Daily Papers' note
score 4