The engineer's race car and its driver (episode 14)

From episode 14 of A Million Tries ("My AI Designed a Race Car With No Front Wing"). Two AIs: the engineer (evolution) sets 13 dials of a four-tyre race car under a $1M price list (engine power and position, front and rear wings, front and rear tyres, wheelbase, drive split, brake balance, roll balance, gearing, ride height, springs), and the driver (PPO) learns to drive whatever it builds, on circuits it has never seen.

This is the engineer's final car (eng-v4-2.json, $972k) and the driver trained on it (100M steps from the general driver, on designs within 15% of it). The car: no front wing at all, a near-full rear wing, the engine at the back, wide rear tyres (405 mm, fronts 275), 510 kW, rear-wheel drive, a low floor (49 mm) on springs stiff enough to stay above the floor's stall, the longest gearing (380 km/h).

The final race (60 s from a standing start, each car with its own driver):

car	cost	56 test circuits	24 never-used circuits
the grid's best (a power x wings sweep)	$979k	3,390 m	3,418 m
the engineer's car (this one)	$972k	3,554 m (2% crashes)	3,602 m
an F1-like car (stiff springs, over budget)	$1,365k	3,545 m	3,593 m

The engineer's car beats the grid's best by +164 m +- 34 (55 of 56 circuits; never-used 24/24) and ties a car that costs $393k more (-9 +- 38). With one shared driver (the general one) the design alone is +3.0% over the grid's best (+95 m), so more than half of the gain is the car itself.

The game: a four-tyre car on real-scale circuits (grip and slip per tyre, downforce from wings and a floor that stalls near the road, tyre heat and wear, fuel, hills and banking, grass that's slow and dirties the tyres). The physics is road-tested against a 911 and an F1 car. Code at the tag episode-14: https://github.com/backstab319/a-million-tries/tree/episode-14
Sees (82 numbers): 13 distance lasers (up to 200 m), the road's bends, climbs and banking at 10 points ahead (10-320 m), its motion (speeds, yaw, g-forces, place on the road, heading, steering, pedal), each tyre's grip, sliding, temperature and wear, the fuel left, and the car's 13 dials.
Does: steering and pedal (throttle or brake), 20 times a second. Brain: 82 -> 256 -> 256 (tanh) -> 2.
Files: model.safetensors holds the weights (PyTorch names; ailearns.agents.ppo_car builds the same network), config.json the training and game settings.

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

Loading…

The engineer's race car and its driver (episode 14)

The final race (60 s from a standing start, each car with its own driver):

car	cost	56 test circuits	24 never-used circuits
the grid's best (a power x wings sweep)	$979k	3,390 m	3,418 m
the engineer's car (this one)	$972k	3,554 m (2% crashes)	3,602 m
an F1-like car (stiff springs, over budget)	$1,365k	3,545 m	3,593 m

The game: a four-tyre car on real-scale circuits (grip and slip per tyre, downforce from wings and a floor that stalls near the road, tyre heat and wear, fuel, hills and banking, grass that's slow and dirties the tyres). The physics is road-tested against a 911 and an F1 car. Code at the tag episode-14: https://github.com/backstab319/a-million-tries/tree/episode-14

Sees (82 numbers): 13 distance lasers (up to 200 m), the road's bends, climbs and banking at 10 points ahead (10-320 m), its motion (speeds, yaw, g-forces, place on the road, heading, steering, pedal), each tyre's grip, sliding, temperature and wear, the fuel left, and the car's 13 dials.

Does: steering and pedal (throttle or brake), 20 times a second. Brain: 82 -> 256 -> 256 (tanh) -> 2.

Files: model.safetensors holds the weights (PyTorch names; ailearns.agents.ppo_car builds the same network), config.json the training and game settings.

amilliontries/race-engineer-car-ep14

The engineer's race car and its driver (episode 14)

The engineer's race car and its driver (episode 14)