Play2Perfect: What Matters in Dexterous Play Pretraining for Precise Assembly?
arXiv:2606.26428v4 Announce Type: replace-cross Abstract: Multi-fingered robots promise the speed and dexterity of human hands, yet challenging problems such as precise assembly have remained out of reach. These tasks are contact-rich, making data collection for imitation learning difficult, and sparse-reward, making direct exploration with reinforcement learning (RL) intractable. Consequently, prior work has made progress by structuring the problem with specialized grippers, tool attachments,…