Goal is to create a racing agent w/ reinforcement learning in a Godot Env
| Human Driven, 1x speed | RL Racing Agent, 2x speed |
|---|---|
 |
 |
Godot handles the physics simulation, gdrl handles the RL + bridge between simulator and gym API
View experriments/journal.md
1. Pull this project
2. Download Godot 4.6.2
3. Import godot_projects/racing-env-v-1/project.godot using the Godot engine import menu
4. Delete the Sync node in the Game scene
5. Hit Run in the editor
1. Run train.py then run the game
- alternatively, export the game as an .exe and pass the path as a CLI arg for physics speedup
2. Once train is done, a `racer_ppo.zip will be made`, run inference.py to deploy your model onto the track
Environment
- facelift: lighting, meshes, level scenery
- tune raycast sensors, adjust spacing and amount
- currently set to 100 meters max, heuristic march then bin search to get distance of car to road
RL
- model selection, currently using gdrl defaults
- reward function creation
- waypoints, turning penalty, gas penalty, time
Waypoint generation is a mostly automated process to help expedite map creation
Paths are provided as (x,y,z) tuples in path_points/*.txt, and tools/generate_path.gd in the project turns them into a path3D + a CSGPolygon3D to provide the track w/ a mesh
- non-parsable lines are skipped (lets us add comments)
- the attached polygon can then be baked into a mesh (for human eyeballs) and a collision shape to keep track of whether the car's on the road
- waypoint creation is done by sampling the curve at even intervals, first waypoint goes to first point in .txt