... I do think "it's unstable" is kind of a non-answer though. Lots of control problems are unstable, right? For example here's a successful project from 16 ye

... I do think "it's unstable" is kind of a non-answer though. Lots of control problems are unstable, right? For example here's a successful project from 16 years go where Andrew Ng, Pieter Abbeel etc used apprenticeship learning to teach a controller to fly RC helicopters through complex stunts, and they discuss how these flight patterns have to deal with instability, and the controller is eventually doing things very fluidly which humans would struggle to do. https://www.youtube.com/watch?v=M-QUkgk3HyE https://www.youtube.com/watch?v=M-QUkgk3HyE http://heli.stanford.edu/papers/iser04-invertedflight.pdf http://heli.stanford.edu/papers/iser04-invertedflight.pdf So what about bipedal walking is harder than other learned control problems that RL can master (and could handle even on previous generations of tech)?