[RLlib] Fix Unity3D built-in examples action bounds from -inf/inf to -1.0/1.0.

[sven1977]

Fix Unity3D built-in examples action bounds from -inf/inf to -1.0/1.0.

If the action space is -inf/inf, unsquashing (e.g. from PPO's normalized action range of roughly -1.0/1.0) will not happen and the env may complain. Unity3D envs usually use action bounds of -1.0/1.0 as indicated by the hard-coded clipping code in Unity's ML-Agents, doing something like:

action = clip(action, -3.0, 3.0) / 3.0 -> always between -1.0 and 1.0.

Why are these changes needed?

#21109

Related issue number

Issue #21109

Closes #21109

Checks

• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[avnishn]

Does unity automatically do space scaling? Seems fine if they do, just so I'm understanding this correctly. I guess another way of asking my question is: what are the bounds of the underlying space on the unity side? Do we define them however we want to, or does unity?

[sven1977]

Unity (the env) does.