Booster recovery controller from last night. Sim design, training and deployment on hardware took < 1 day. With Qiayuan Liao

@qiayuanliao the solution to the sim2real gap turns out to be not tuning the simulation to reality but rather building the hardware that behaves like the simulation :-)

Study Online with Acellus Gold, Designed to Accelerate Learning & Motivate K-12 Student Achievement.

@qiayuanliao it's really impressive how quickly it stands up

@qiayuanliao Wow I really thought the sim2real gap would cause major issues, but pleasantly surprised that the real robot is able to do the crazy movements it does in sim. Great work!

@qiayuanliao *there’s no way the robot can do that in real* oh 😮

@qiayuanliao Does this type of recovery have an negative impact on motors/casing? Wonder if there's any benefit to making it recover more naturally (humanly), or that's just our personal preference based on what we'd expect.

@qiayuanliao I didn't do any regularization, the goal was to do a dry run with a very simple reward and minimal DR.

@qiayuanliao That recovery on the real robot looks unreal!! 🤯 Was this pure sim to real transfer? If so there hardly seems to be any gap between sim and real! Amazing job! ❤️

@qiayuanliao Yes indeed! Actually, we *only* randomized foot friction during training and it transferred pretty flawlessly.

@qiayuanliao It's pretty cool how it stands up usually when i see behaviours like this in training runs in simulation i think this wouldn't work in real life. But it actually does crazy.

@qiayuanliao Woah that’s insane!