3 Comments
User's avatar
Roy Xing's avatar

Great article! Tangentially, this article is the first time I really thought of the motion imitation learning methods of copying humans mocap data as biomimicry.

Avik De's avatar

Yes, exactly! It isn't as consequential for quasistatic behavior, but with dynamic behaviors or with severe constraints, it will begin to matter more.

Roy Xing's avatar

Yeah, one minor experiment I ran was having AMP use dog data vs MPC data. MPC data was way more robust and energy efficient, which in hindsight is super obvious since the behaviors of MPC is meant for the same embodiment. However, I do admit that the dog behavior "looked" way nicer, so I wonder how much bias we as a community now have for what looks nice over solid measurable quantities for robot motion!