> such that they can use the right set of sensors in the right environmental conditions

Because this part is really hard, and that's why Tesla abandoned the fusion approach. You cannot possibly foresee all the conditions in which LIDAR or any active sensor will malfunction/return wrong data/return data that's only slightly off for that ONE specific time. And even if it doesn't, you need to trust it to not return noise. And when it does return noise, how do you classify it as noise?

Cameras are passive sensors - they get whatever light comes in and turn it into an image. Camera is capturing shapes that make sense to the neural nets: it's working. See all black/white/red/cannot see any shapes? Camera is not working, exclude it from the currently used set of sensors or weigh it less when applying decisions, because it's returning no signal (and yes, neural nets have their own set of problems).

EDIT: cameras also provide more continuous context: if 1 pixel is off, is clearly bright red in a mostly-green scene where no poles can be identified, the neural net will average it out and discard it as noise. If 1 pixel says "object" in LIDAR, do you trust it to be correct? Perhaps the ray just hit a bird or a fly, but you only see a point, it's a lossy summary of the information you need.

They could in theory. If they put at least as much emphasis on the AI side as Tesla does. Or if someone else cracked vehicle AI wide open and left it open for them to copy, and then they did exactly that, and found a way to bolt on their extra sensors in a useful fashion while at it.

As is, Waymo's playing it smarter than Cruise did, but they're not all in on AI yet. So I don't expect them to "leapfrog Tesla" in that dimension - and it's the key dimension to self-driving.

I thought about this and I think it boils to how the model is trained.

Tesla trains it models from actual drivers purely based on (input) Vision and (output) actuators - Brake, Steering, Accelerators.

Human output is based on what they and the camera sees. So, it's a 1:1 match.

If Waymo were to do that, it'll muddle the training set. The Lidar input may override camera input.

I always struggled when Musk mentioned Lidar will make it ambiguous. It didn't make any sense to me why having a secondary failback sensor messes things. But, if you put it in the training data context, it absolutely makes sense.

Because they don't have a fleet of millions of people labeling the data for them and paying for the privilege of doing so. Waymo has about 3700 vehicles. Tesla has millions. Waymo only operates in known environments and collects a very limited range of data. Tesla collects data everywhere that people drive their cars.

The main reason Tesla's don't have LIDAR is hardware cost and maintenance cost, not improved safety.

I got downvoted for saying this last time the topic came up but constraints focus a project. It’s best to start work with as few variables as possible, and only add new ones when absolutely necessary.

I'm working on a similar problem in computer vision and we're quickly approaching the point where our pure vision work is better than our Lidar supported track because we've had to deal with the constraints instead of having a crutch to lean on.