TL;DR: Real Robot Challenge Phase 2: No ML Required
Abstract: In Phase 1, we presented an approach to rigid-body manipulation based on carefully-designed motion primitives. We now demonstrate in Phase 2 that these primitives combined with a more robust state machine can reasonably complete all 4 tasks. For Phase 3, we propose using model-based and time-constrained (e.g. contextual bandit) RL to select and optimize primitives based on the environment. In general, our approach emphasizes the use of ML machinery only when classical approaches fail. Video Attachment Playlist: https://www.youtube.com/playlist?list=PLEYg4qhK8iUaXVb1ij18pVwnzeowMCpRc This playlist includes one video of a successful attempt for each difficulty level.