C++20/Rust Rocket League RL trainer built around RocketSim and LibTorch.
The main baseline here is normal Soccar 2v2 self-play with PPO/GAE. It has batched inference, checkpoint resume/versioning, old-policy sampling and a separate evaluator.
git clone --recurse-submodules https://github.com/Speedstu/RocketLearnNative.git
cd RocketLearnNativeAlready cloned without submodules:
git submodule update --init --recursiveBUILD.bat
START_TRAINING.batBUILD_EVALUATOR.bat
START_EVALUATOR.batrocket_learn_native is the trainer and rocket_learn_evaluator is the deterministic 2v2 evaluator. There is also a small Rust supervisor for restart/log handling.
Most experiment-specific launchers live next to the main scripts. They are there because I use this repo as a working training tree, not as a polished library.