HOC2R releases method for extracting hand-object contact trajectories

Hongkun Yu and other authors propose and release HOC2R, a method for extracting hand-object contact trajectories from human videos for robot motion transfer. It combines frozen DINOv3 and SAM 2.1 features with learning heads to perform hand-object association and contact localization. On a 30-task-point annotation benchmark, the authors claim that count-aware automation raises contact-count accuracy from 58.1% to 71.4% and reduces average localization distance from 19.81 pixels to 10.46 pixels, compared with an adapted VRB2 baseline. On 7 unseen RGB-D rollouts, the five-task fixed-two checkpoint reduces average localization distance from 67.67 pixels to 27.02 pixels. In 6 robot evaluation sequences, the 30-task fixed-two checkpoint produces 4 feasible trajectories and successfully completes 3 tasks; the authors claim each adapted baseline produces 2 feasible trajectories and 1 task success. The authors also claim the method is competitive in the evaluated single-hand scenarios, is robust to hand color changes, and performs strongest in the evaluated multi-hand scenarios.

AI 根据报道生成 · 2 days ago更新