Comment by fastball | Hacker News Reader

Comment by fastball | Hacker News Reader

Parent Full thread

How is that not an argument?

A human can drive in the snow, and has a sensing package which effectively consists of two cameras.

The hard part (when you only have two cameras) is that the processor needs to be exceedingly sophisticated, e.g. a human brain.

People like Waymo have decided the best way to decrease the brain requirement is to add more sophisticated sensing that humans are fundamentally incapable of.

Tesla's strategy is more to just build a better digital driving brain with machine learning + lots of miles, because arguably all sensing mechanisms have their shortcomings. That being said, they still have a lot more than two cameras in their FSD package – just not LIDAR. They still have radar and other things as well.

Only time will tell which approach is better. Probably in the end the best approach will be to combine the two: a brain that is sophisticated enough it can drive with just cameras, but you throw LIDAR and such in their too to go the extra mile.

> A human can drive in the snow, and has a sensing package which effectively consists of two cameras.

Two cameras, and a huge amount of mostly not understood wetware interpreting their signals. Our current approaches to ML may be a local maxima. The true path to AGI be somewhere as yet undiscovered. Lidar certainly seems like a good tool to use given our current ML capabilities.

Our current ML capabilities involve throwing as much training data at the problem as humanly possible.

No LIDAR means more training data, as has been made clear by the amount of miles driven by Waymo vehicles vs Teslas in Autopilot.

it's a strange philosophy. Isn't the point of building machines that they can do things that humans can't? It's like building drones for warfare and turning the heat-seeking off because humans don't have that capacity.

A generally intelligent driving system is going to benefit from better sensory inputs just like a human would, it's not like that makes the system dumber. The whole dichotomy between sensory inputs and processing seems strange to begin with, there's not necessary a clear line, or a reason why one should be prioritized over the other.

We are not talking about generally intelligent driving systems though.

You have to realize that most SOTA self-driving systems are using a lot of NNs. NNs in turn want lots of training data. That training is not fully generic – in other words, if you train a NN using only cameras, giving LIDAR to that trained network (i.e. for inference) will not really do anything.

Having large LIDAR arrays on the top of vehicles is expensive and slow to deploy, which directly translates to being able to generate less training data.

So if your goal is to have the best possible self-driving vehicles 100 years from now, absolutely you should be using as much sophisticated sensing equipment you can get your hands on, and slowly train your nets for the next century. If your goal is to get out something that is better than a human driver as fast as possible, then it is not clear that requiring all vehicles to have LIDAR on them is the best way to go.

Not necessarily. Usually machines are built to do mundane tasks that humans have to do. In most cases they are built to replace what a human can do.

The argument for having a lidar vs not having a lidar is different. They chose to go without lidar because it should be cheaper and faster to get started, and the fact that people are capable of doing driving without lidar makes it seem possible that a machine could do it without lidar.

Reply on news.ycombinator.com