SELECTED PROJECTReinforcement Learning

TrackMania 2020 Reinforcement Learning Agent

A Soft Actor-Critic agent connecting reverse-engineered TrackMania telemetry to real-time control.

  • Rust
  • Python
  • PyTorch

Objective

Build a reinforcement learning control loop for TrackMania 2020 using observations extracted directly from the game.

System

Reverse-engineered game telemetry to expose virtual LIDAR observations and used dynamic memory hooks to send real-time controls. The implementation combines Rust-based game integration with a Soft Actor-Critic agent trained and evaluated through Python and PyTorch.

Outcome

Produced an end-to-end loop that connects live game state, learned policy inference, and low-level control without claiming an unverified lap-time or win-rate result.