[cropsieboss]

The thing is, for car driving you need combinatorial search combined with a good model (statistical) of the world. Research I've seen does not incorporate any kind of meaningful search.

Company that has all the roads and surroundings scanned and knows exactly where the car is at all times is at a huge advantage.

Partially observed world gets your head hurting. Not modelling other cars practically makes everything useless.

If you can't hallucinate everything around you as close as possible, the problem gets very very difficult.

I'm pretty sure there are companies close to solving the problem, as they have more data and some also know practically every road in the modern world.

Modelling and tracking other cars is a bigger issue, including pedestrians.

[dclowd9901]

> Modelling and tracking other cars is a bigger issue

Aren't there very reliable ways to model other cars? Something as big as a car can't move very unpredictably in a millisecond-to-millisecond basis. And at the end of the day, all you need to do is be able to model the behavior of your car.

If car in front of me starts to swerve over on a collision course with my car, my car can note the probability of a collision a long (computer-long) time before it really needs to do anything about it. Knowing full-well its own braking, swerving and acceleration capabilities at any given speed, it should be able to know when it _must_ take action to avoid the actions of another, and it doesn't even have to rely on reaction time.

No, I think the real problem here is for cars to do things like understand _context_ of a situation, like, for instance, how can it tell that there's a bus in front of me turning around that's going to be clear of my way in a moment vs. a tree that's been felled? That's something that computers are and will probably continue to be terrible at until they can closely mimic the capabilities of the human brain.

[bitL]

> Modelling and tracking other cars is a bigger issue, including pedestrians.

Unscented Kalman filters and particle filters are used for that - they are pretty good if you sample environment with >10Hz rate. Most of what you said has been taken car of pretty well, those aren't really hot issues anymore. Also, the thinking these days is that humans (even mediocre) can do a good job driving using eyes only and more robust camera solutions based on deep learning are here. Their problem is inference speed as HW is not fast enough to allow FHD or 4K input feed to be processed realtime, so imagery has to be downscaled significantly, leading to errors. Similarly, tracking other cars only depends on how big supercomputer can you fit into your trunk.