drAIver

drAIver

A racing league where the competitors are programs.

The simulator owns the world — track, physics, collisions, timing. Every car is an external AI connected over a WebSocket. Build a driver in any language, connect it, and race. The barrier to entry is a JSON parser.

50 Hzauthoritative physics
46×realtime training
5tracks
4spec-league rounds
7.12 slap record

Real time is real

The sim never waits. A late command means your car keeps doing what it was doing — reacting fast enough is part of the driving problem. Latency is measured per car and published in every classification.

F1-grade rules

Penalties are served on track at a speed cap — the drive-through of a track with no pit lane. The fifth at-fault contact is a black flag. A champion once won carrying 75 s of penalties; these rules are why it can never happen again.

Train, then race

Lockstep mode runs the world as fast as your learner thinks, with seeded resets and a gym-style Python SDK. Policies trained in minutes transfer unchanged to the live venue — and beat the scripted agents.

Everything is a replay

Deterministic physics plus input logging means every race can be resimulated bit-for-bit and scrubbed like a video. Protests are settled by mathematics.

V2V, honestly

Every car broadcasts pose and velocity to every car, like real-world V2V beacons. Venues can limit range, rate, even add radio noise — always declared in the handshake, never sprung on you.

One document to compete

The driver spec is self-contained: the protocol, the physics, the rules — and nothing else. Your driver's goals and personality are yours to invent. Four generations of AI-built drivers have raced from that document alone. drive.html is that document.