A custom environment made using the Godot Engine and written in C# used to train reinforcement learning models to drive a car on a top-down virtual track with semi-realistic physics. Utilizes the Godot RL Agents framework to train agents using the Stable-Baselines3 implementation of the Proximal Policy Optimization algorithm.

Full Demo Video

Full Demo Video

Inside of a Python virtual environment, install the Godot RL Agents package:

pip install godot-rl

Download the latest release and run the following command to begin training after activating the virtual environment:

python "sb3.py" --env_path=rl-racer.exe --save_model_path=models\model --n_parallel=4 --speedup=16 --save_checkpoint_frequency=50_000

Or start running inference on the pre-trained model:

python "sb3.py" --env_path=rl-racer.exe --resume_model_path=models\model-3 --viz --inference

• Added training of multiple agents
• Added multi-target camera tracking

• Added an angle limit to reach fail state
• Added a small penalty for driving off-track
• Increased off-track friction
• Slightly increased slip speed threshold

• Time-tracking and checkpoints
• Human controlled vs. Agent(s) controlled
• Improved collision avoidance
• Linux and maybe macOS binaries