The day in conversation for 2026-10-19

The discussion: Dev: I’m still stuck on how much weight to give to KODA versus a purely supervised imitation learning setup. Forty-five percentage points improvement is significant, but is that improvement robust enough across different robot morphologies? Rosa: I think the G1 results are promising because they prove that synthesizing whole-body data from just egocentric RGB input isn't just theoretical magic. It shows the alignment mechanism actually captures the necessary physics constraints better than simple end-effector tracking. That fixes a huge problem we’ve had since last week when our models kept drifting away from valid kinematic paths during complex gaits. Taro: But look at MoCaP, Dev. If KODA is focused on aligning demonstrations to whole-body action synthesis, MoCaP seems to be focusing on the interpretation of passive cues—spatial memory for navigation and tool use. That feels like a much more general problem for long-horizon tasks than just nailing a specific locomotion trajectory. Dev: True, MoCaP tackles the *why* behind the