This browser could not start WebGL, so there is no city to drive through.
watch a car learn to see, decide and drive
Think of a driving school with one examiner. Every driver, whether it is a total beginner, a student with an instructor, or a graduate, sits the same test and is marked on the same report card. That is the only fair way to compare them.
Knows nothing, not even that the right pedal makes it go. It tries things and the examiner's marks come back as reward and punishment. Like a toddler learning to walk, it falls a lot first. It works up through five levels, from an empty straight road to rush hour.
Has an instructor with a second brake pedal. It watches, guesses, and the instructor grabs the wheel when it sees where the student is heading: out of the lane, off the road, or too fast towards something. It hands back only once the student shows, for a whole second, that it would do the right thing. Only the good stretches of the instructor's driving are kept as lessons.
A network that already studied for a few simulated hours. It does not learn while you watch: it is the standard to aim for.
The hand-written planner. It is the ceiling for a student, and it makes mistakes too. A student that copies it cannot beat it.
Green dots: this driver's test. Every few minutes of practice, a frozen copy of the driver drives the same 150 seconds in the same city with the same traffic and the same three surprises (a crate, a jaywalker, a stalled car). Nothing is learned during the test and nobody grabs the wheel. A rising line can only mean it really drives better.
Dashed lines: the yardsticks. White is a driver that never learned anything (day one), orange is the instructor, blue is the pre-taught graduate. The gap between white and green is what learning has bought so far.
L2, L3 marks (ground zero only): the moment it passed a level. A new level usually knocks the score down for a while, because the world just got harder.
Eleven marks, out of 100: safety, lane keeping, kerbs and edges, being in the right lane for the route, smoothness, indicators, speed, stopping at lights, following distance, parking and self-reliance. The overall score weights safety heaviest. The same marks are what the ground-zero driver is rewarded and punished with, and what filters the instructor's lessons for the student, so what you read there is what it is optimising.
A small city, live traffic and pedestrians, and one self-driving car. Everything the car does, it does from its sensors: it never gets to peek at the simulator's answers.
Vision reads paint and signal colours but guesses depth and goes blind at night and in fog. LiDAR measures exact 3D shape with real laser rays but can't read a red light. Fusion cross-checks both.
LiDAR points are clustered into boxes; camera detections are matched to them; a tracker keeps identities and velocities. Red rings are things that were really there and the car missed.
Seven candidate paths fan out. Each is scored against where everyone is predicted to be. The green one wins; red ones run into something.
Pick how much the driver knows. Ground zero starts with nothing and learns by trial and error, marked on everything an examiner would mark. The Student has an instructor: it watches, tries, and gets corrected. The Graduate is already trained. The Is it learning? card shows a fixed driving test and a report card, so you can see real progress, not just luck.
Space pause · 123 sensors · V camera · H hazard ahead · QE turn a hazard you are placing · N new city · [] speed · JL change lane · P pull over · T turn round · X hold / go · drag to look around · right-drag or shift-drag to pan · wheel or pinch to zoom · double-click to reset the view · click the map to send the car somewhere · on the Hazards tab, click a card then the road, or drag it straight onto the street
The camera "detector" is a physical error model (occlusion, contrast, range, weather) applied to what is really in front of it, not a trained vision network. LiDAR is real ray casting. The driving network is real and trains live in your browser.