HumanPlus is a phenomenal project. I’m deeply impressed by Tony, Chelsea, and Mark.
These researchers built their own humanoid and have been training it with imitation to do fine grained tasks. It’s a natural extension of ACT & Aloha.
All this project is, is a basic pose estimation model coupled with ACT. They made a nicer interface for collecting the data — makes complete sense this was the exact direction I was headed in.
The next evolution is universal translation. That is, creating a standard way of doing this sort of training but not coupled 1:1 with a human skeleton.
I was excited since this validated my own concept. All lights are go on Droids. I have everything I need. All that remains is (1) making this toolchain accessible and (2) adapting the interface to many arbitrary form factors.