What you're looking at
This is the MuJoCo model from the study, compiled to WebAssembly and stepped at 500 Hz in your browser,
with the same torque-limited jaw loop, pad contacts and crack rule. The policies are the trained networks,
run live at 30 Hz. Nothing is pre-recorded.
Training happens offline (the 80 policies behind the gain chart took about 2.5 hours on 10 cores), so
the choices above pick from policies already trained by the code in the repo, each the seed closest
to its condition's median.