Ann: Brainchip Race Car Video Presentation, page-38

"This has nothing to do with the quality of graphics on display."


Lets ignore the graphics for a moment.

It is possible to perform a simulation like this with no real learning at all. You simply need a "random walk" generator and have negative feedback.

The "car" [actually just a block of sprite pixels] simply has to head in the general direction of "ahead" with random right or left turns of less than 90 degrees. If it hits a wall it gets a signal to turn a few degree in the opposite direction and keep moving forward. It can easily keep track of where these bumps occur [a simple time/distance calculation] and turn away before hitting the wall on the next lap. With each lap it turns more quickly so it becomes faster. However the car has not learned anything except bouncing a bit before it actually hits the wall.

The above simulation idea looks impressive but it is completely useless because the "car" can't use that "knowledge" to negotiate a different track. It literally has to relearn everything from scratch every time the track changes slightly [because it is literally just bouncing off the wall.]

The Google car on the other hand literally reads traffic signs, identifies pedestrians and is aware of other vehicles and complex road behaviour.