Beschrijving
Learning dexterous manipulation typically requires hundreds of demonstrations, making data collection costly and hard to scale. We propose a three-stage framework learning from a single human video. First, hand and object poses are extracted from the video. Second, a motion prior policy is trained via imitation learning. Third, residual RL fine-tunes the policy, and diffusion distillation compresses it for real-time deployment. The approach achieves strong performance on dexterous manipulation tasks with minimal human data.
Website: share.transistor.fm

