The race-day driver (episode 15)

From episode 15 of A Million Tries ("My AI Race Car Refused to Make a Pit Stop"). One brain drives every car in an 8-car race: contact, damage through the physics, dirty air, soft/medium/hard tyres that wear out (scaled to a Grand Prix's 305 km) and fail when worn through, a fuel limit with no refuelling, and a pit lane with an automatic 80 km/h limiter. The team calls it in; it has to turn in, brake and stop on its mark.

How it got here (devlog steps 8-9): the episode 14 driver, then PPO in a race school (one-minute stretches of real races, the team paying back a called stop, a pit coach): ~60,000 calls, 0 stops. Then imitation of a hand-written pit driver (turn-ins 2 -> 146 of 256 called cars on unseen circuits), then evolution on race results (finish, fast; not finishing costs; no coach): whole races, then Grand Prix tyre wear. This is that last run, generation 50.

Under the final rules (marshals clear a stopped car in 10 s, an empty tank retires the car), 256 car-races in mixed fields of random designs on test circuits, GP tyres, the team calling when the worst tyre passes 60%:

• Sees (115 numbers): episode 14's 82 (lasers that now stop on other cars too, the road ahead, its motion, tyres, fuel, the car's dials) plus the two nearest cars ahead and behind, what's broken on its own car, the tyre compound, the pit call and a sharp sense of the metres to its box.
• Does: steering, pedal and reverse, 20 times a second. Brain: 115 -> 256 -> 256 (tanh) -> 3.
• Code at the tag episode-15: https://github.com/backstab319/a-million-tries/tree/episode-15 (games/race_field.py, agents/es_field.py; ailearns.agents.ppo_car builds the same network).

From the YouTube channel A Million Tries: https://www.youtube.com/@amilliontries