Self-driving car
A small neural network drives a curvy highway through traffic. It keeps to its lane, decides when to brake and when to overtake, and signals before it moves over. It learned all of that by evolution, with no examples to copy, only distance covered and a great many crashed cars. Watch it, evolve your own, or take the wheel.
1
0.00km
0km/h
0s
- 00–04
- lane to the left
- 15
- speed
- 05–09
- own lane
- 16
- curve ahead
- 10–14
- lane to the right
- 17–18
- lane change
←→change lane ↑raise set speed ↓brake Ragain
Pretrained
Highway
Three lanes
Anatomy
Live from the car in focus · gaps on a square-root scale
How it learns
Nobody shows the network how to drive. A population of random networks takes to the road, and the ones that get furthest become the parents of the next generation.
- 01Sixty cars with random weights start on the same road. The planner keeps each one in its lane, so even these drive; what they lack is any idea of when to brake or which lane to be in.
- 02Each drives until it crashes, falls behind a line moving at 18 km/h, or the round is up. Rounds start at 45 s and grow by half each time a car is still driving at the end, up to five minutes.
- 03Its score is the distance covered, minus 300 m for going out early.
- 04The best four carry over. The rest are children of tournament winners, crossed and mutated.
- 05Overtaking takes far longer to find than sixty cars and a few minutes allow. From pretrained starts the field from mutations of the shipped brain instead, which changes lanes from the first round.
Who decides what
The network chooses the pedal and the lane. A planner does the rest, the way a driver-assist system would: it blinks for 0.7 s, waits while a car is level with it in the target lane, moves over on a 2.6 s S-curve and steers along it with pure pursuit. The blue ribbon is that plan. Whether to overtake, and whether the gap is worth it, is the network's call alone. Its score loses a little for every second it cruises left of a free lane, so it learned to move back right after passing.
The trained driver
The brain in Watch came out of a longer run, headless in Node, on five-minute episodes in full traffic. These are its numbers on 30 roads it never saw. Traffic drives around it like people would: it keeps its distance, signals and overtakes, but it never makes room for a car that cuts in.