Yunfei Xie Introduces RoboTok for Robot Demonstration Retrieval

Yunfei Xie announced on X that RoboTok treats web videos as an internet-scale source of human demonstrations for dexterous robot learning. According to the post, given a query video, RoboTok finds other clips with similar hand motions by comparing canonicalized 3D trajectories in an actor-centered frame and avoids reliance on labels or visual appearance. The post states experiments showed it retrieved more relevant demonstrations than prior methods, and simulated robots trained on the retrieved videos completed tasks more successfully. Xie credited collaborators and linked the project site, arXiv paper, code, and public models.