This project features an AI agent designed to play Trackmania, using computer vision and reinforcement learning techniques.
Note: This project drew heavy inspiration from Linesight.
I chose a QR-DQN because it handles unpredictability well and can make riskier choices when the strategy calls for it.
- 120x90 grayscale frame of the game (1, 90, 120)
- Current speed (1,)
- Yaw (1,)
- Velocity (3,)
The model outputs an integer from 0 to 5, each representing a discrete action:
| Value | Action | Controls |
|---|---|---|
| 0 | Forward | left: False, right: False, accelerate: True, brake: False |
| 1 | Forward left | left: True, right: False, accelerate: True, brake: False |
| 2 | Forward right | left: False, right: True, accelerate: True, brake: False |
| 3 | Nothing | left: False, right: False, accelerate: False, brake: False |
| 4 | Nothing left | left: True, right: False, accelerate: False, brake: False |
| 5 | Nothing right | left: False, right: True, accelerate: False, brake: False |
This project is licensed under the MIT License.