[ckemere]

Training data of task completion. See, e.g., robots doing backflips. Presumably there’s an optimal robot for gymnastics but if you start with humanoid form you can train based on many videos of human movement. The alternative - world model sim with physics and loss functions- perhaps ends up being too unconstrained when you add in optimization of the robot form…

[gowld]

AI watching videos of humans seems an incredibly inefficient way to solve acrobatic balance. It is just physics and engineering. The hard part is in the materials and assemblages for enabling complex and subtle movements and the fine motor control, not knowing what the motions are.