03 — private · research
Trackmania General AI
A general Trackmania driver: pixels in, steering out, on tracks it has never seen.
python · pytorch · rl · tminterface
~10k
tracks in the training pool
50
attempts to gold, the benchmark
0
game‑engine inputs
the idea
Most Trackmania agents learn one map. This one trains with reinforcement learning across a pool of roughly 10,000 tracks so the policy has to learn driving itself — braking points, drift, walls — rather than a memorised line. The only input is the rendered frame; no game‑engine state, no track geometry.
the benchmark
Gold@50: reach the gold medal within 50 attempts on held‑out maps. The baseline and infrastructure come from Linesight; the agent code builds on top of it.
the machine
Training runs on the Legion 7 — RTX 3080 Laptop, Ryzen 9 5900HX, CachyOS — reached over the local network and Tailnet. The repository is private while the work is in progress.