Event

CARD: A Contact-Anchored Retargeting Framework Without Simulation or Object Models

1 reports · 1 sources

Cunxi Dai and other authors released a paper proposing the CARD framework: it does not rely on simulators or object models, and by explicitly preserving contact force objectives, it retargets human demonstrations into robot actions, and uniformly handles video, motion capture, and tactile glove data. To maintain contact objectives during open-loop hardware replay, CARD also introduces a tactile tracking controller. The authors claim that on video, glove, and motion capture data and the GRAB dataset, CARD maintains contact on 43%–78% of contact frames, whereas pure kinematic and object-aware retargeting methods are below 10% and 22.5%, respectively; on the GRAB dataset, after integrating this method into physics-based retargeting, the strict task success rate increased from 12% to 32%. These are all results reported in the paper.