Event

AdaDexTrack proposes a language-conditioned dexterous manipulation tracking framework

1 reports · 1 sources

Jianibieke Adalibieke and other authors released a paper proposing AdaDexTrack, a "modulator-in-the-loop" framework for language-conditioned dexterous manipulation tracking. Its approach adds three types of feedback correction outside the tracker—reference modulation, object latent modulation, and position target modulation—to alleviate synthetic reference noise, embodiment bias from human-to-hand retargeting, and error accumulation in fixed-reference tracking. The authors state that the tracker is learned via large-scale "expert-to-generalist" distillation on a language-conditioned hand-object trajectory corpus, while the modulators are trained with reinforcement learning under the same task objective, tightly coupling the two. The authors also state that zero-shot sim-to-real transfer was achieved on real hardware, and that adding the modulators yields significant improvements compared with using only the tracker; these results are all claims made by the paper's authors.