Merge pull request #5 from anjo7778/codex/integrate-rl-gym-for-reinforcement-learning

anjo7778 · web-flow · commit b8fbae28ac26 · 2025-12-03T19:02:32.000-05:00
Update torch requirement for newer Python
diff --git a/README.md b/README.md
@@ -5,7 +5,20 @@ RLGym example bot for the RLBot framework, based on the official RLBotPythonExam
 
 This bot runs the Actor class in `src/actor.py`, you're expected to replace that with the output of your model
 
-By default we use DefaultObs from RLGym, AdvancedObs is also available in this project. 
+By default we use DefaultObs from RLGym, AdvancedObs is also available in this project.
+
+## Using RLGym resources
+
+The builders and parsers in this project come from [RLGym](https://rlgym.org/). Their documentation explains how
+observations and actions are structured during training, which directly maps to the `DefaultObs`/`AdvancedObs` and
+`DefaultAction` helpers shipped here.
+
+- Review the observation/action spaces on the RLGym site to choose the builder that matches your trained agent.
+- Keep the `tick_skip` in `src/bot.py` aligned with the value you used when training through RLGym.
+- When loading a trained policy in `src/agent.py`, ensure the state formatting mirrors the RLGym observation builder and
+  parser combination you selected.
+
+Point your teammates to https://rlgym.org/ if they need an overview of how the environment works before swapping in a model.
 
 You can also provide your own custom ObservationBuilder by copying it over and replacing the `rlgym` imports with `rlgym_compat` (check `src/obs/` for some examples)
 
diff --git a/src/requirements.txt b/src/requirements.txt
@@ -2,7 +2,9 @@
 # You will automatically get updates for all versions starting with "1.".
 rlbot==1.*
 --find-links https://download.pytorch.org/whl/torch_stable.html
-torch==1.9.1+cpu
+# Torch 1.9.1 wheels are no longer published for newer Python versions.
+# Pin to a recent CPU build that is available on PyPI to avoid installation errors.
+torch==2.2.2
 rlgym-compat>=1.1.0
 numpy
 
