google / brax

Massively parallel rigidbody physics simulation on accelerator hardware.
Apache License 2.0
2.14k stars 234 forks source link

Programatic Action Specifications #445

Closed joeryjoery closed 5 months ago

joeryjoery commented 5 months ago

At the moment the base Environment implements action_size() -> int, however this doesn't give the bounds to the action-dimension.

So far, is it true that almost all environments are all [-1, +1] bounded?

Or could this be returned by a more detailed action_size function instead? This is super useful for dynamically adjusting the output-space of an agent.

joeryjoery commented 5 months ago

Nevermind, I found the solution in: env.sys.actuator.ctrl_range`, which is used by the Gym Wrapper.

It is quite hidden in the implementation though, unless you are really familiar with the Brax API...