How kgolemo/f1-racing-line-optimization works
Project type and purpose
• This is an open-source reinforcement-learning (RL) application whose goal is to optimize the racing line of an F1 car.
• Key indicators: Python scripts for environment (car.py, track.py), RL training (ray_training.py), rollouts (ray_rollout.py), evaluation/rendering (render_agents.py, race.py), and saved states (exported_states/*.npy).
File structure
• Core environment modules:
– car.py: vehicle dynamics, state transitions
– track.py: circuit geometry
– utils.py: shared helpers (e.g., state preprocessing)
• Training pipeline:
– ray_training.py: orchestrates RLlib training, defines policy, hyperparameters
– exported_states/: stores checkpoints and rollouts in NumPy format
• Rollout & evaluation:
– ray_rollout.py: uses trained policy to generate trajectories
– render_agents.py, race.py, manual_mode.py, use_agent.py: interactive and batch–mode evaluation and visualization
• Assets & media:
– imgs/, media/: animations and GIFs for README
• Configuration & dependencies:
– No Docker or CI scripts found; Python requirements inferred via imports (NumPy, Ray RLlib, Gym-style interfaces).
README for additional insights
• Confirms RL application for game-style simulation based on a Toward Data Science article.
a. Main components:
1. Environment Layer
– Car Dynamics (car.py)
– Track Geometry (track.py)
– Utilities (utils.py)
2. RL Training Layer
– Ray RLlib Trainer (ray_training.py)
– Hyperparameter Config
3. Data Storage
– Checkpoint Store (exported_states/*.npy)
4. Rollout & Evaluation Layer
– Rollout Engine (ray_rollout.py)
– Interactive/Evaluation Scripts (race.py, manual_mode.py, use_agent.py)
5. Visualization Layer
– Renderer (render_agents.py)
– Media Assets (imgs/, media/)
6. External Services
– Ray Cluster / Local Ray instance
b. Relationships & interactions:
• Trainer ↔ Environment: Trainer samples episodes via Gym-style interaction with car/track modules.
• Trainer → Data Storage: Periodically writes checkpoints (.npy).
• Rollout Engine → Data Storage: Reads checkpoints to generate trajectories.
• Rollout Engine → Visualization: Feeds trajectories to render_agents or race for playback.
• Interactive Scripts → Environment: manual_mode & use_agent hook into the same environment for testing.
c. Architectural patterns & design principles:
• Modular separation: environment, training, rollout/eval, rendering
• External orchestration via Ray: supports distributed training and rollout
d. Key technologies & frameworks:
• Python 3.x
• NumPy for state storage
• Ray RLlib for distributed RL training
• Gym-style environment interfaces
• Matplotlib / custom rendering in render_agents.py
Essential elements
• Don’t omit any script that plays a role in the RL loop.
• Every directory of assets or data storage.
• Both automated (ray_training, ray_rollout) and interactive (manual_mode, use_agent) paths.