They moved to pure vision because they were constrained by radar suppliers because of supply chain issues and the chip shortage, not because of any ML progress - they were actually investigating using a higher resolution imaging radar before the pandemic.

They also have Lidar cars in Fremont that drive around every so often, that doesn't mean they plan to go into every car anytime soon. It'd be short-sighted to not continuously evaluate solutions that previously had constraints (despite what Elon says, they'd add Lidar if it made economic sense and it showed an improvement over camera-only, as their camera detection is pretty accurate these days in FSD).

I’ve seen their LIDAR cars in SF too - if I had to guess they’re gathering ground truth data to train monocular depth models on.

And even really naive integrations of LIDAR will show big improvements over camera only. You can do something as simple as overlay the returns from the most recent LIDAR spin over a camera image as a fourth channel and feed it into your models and most of your depth prediction/spatial predictions will improve dramatically.

No, they're not. The linked post even mentions Softbank being a shareholder. It would be good to know what % of Cruise is owned by GM, but as far as I know it hasn't been made public knowledge. Hope someone can correct me on that.

This is true but there’s a bit more to it. Going vision-only is a big move that requires innovation in a lot of areas and changes huge parts of the stack.

Tesla was already heavily relying on vision before going vision-only.

I don't know who owns them. (see other discussion). I do know that automakers have aggressively partnered on self driving. Once this is ready for mainstream a most automakers will all introduce it because they all have rights to it.

I don't know the contract, it is likely that there is some ordering where the luxury brands can sell it first. It will be rolled down to cheap brands as needed (depending either on market demand, or legal mandates)

The reason camera only won’t happen is because it won’t work. Elon has been fighting reality for years.

How does a human identify obstacles, VRUs [0], and other cars?

0: vulnerable road users, eg pedestrians and bicyclists

> How does a human identify

With the most complex, context-aware, intuitive computer in existence. In addition to eyeballs that are dramatically more capable than any camera Tesla is using.

Not going to comment on the overall lidar vs. pure camera approaches, but I don't think human eyeballs are more capable than a full 360-view array of cameras looking at everything (including blindspots) at once. For once, human eyeballs cannot see in full 360 or view the environment from outside of the car cabin (aka unrestricted by pillars and other things that obstruct the field of view from inside the cabin).

Have you ever used night mode on your phone? It looks nothing like what you are taking in terms of colours, but it shows a lot more detail than the human eye can pick up. Same thing with zoom cameras. Also there are more cameras than eyes, and in better spots.

Have you ever used night mode on your phone while driving on 60mph at night? It's basically sophisticated long exposure that won't work for driving.

How does a bird fly?

For human flight, we borrowed the wings, but made them fixed and added propellers, to cheat the mechanical incompetence.

I wouldn't be shocked that radar/lidar/sonar/whatever sensors are what it takes to cover the incompetence in matching human brain+vision.

Heck, use multiple "brains" and give each veto power on moving the vehicle. Supposing that stopping doesn't kill you, that would at most frequently annoy the driver, and sometimes save his life or someone else's.

Well also the largest animal to ever fly only weighed a few hundred lbs so we’re also just limited by the kind of flight we’re trying to do.

At least today we probably could build a craft that flies by flapping its wings but what would be the point?

Forget birds, insects do it, with vision and puny brains.

The human visual system works more like a high resolution event camera than a frame-based camera. Event cameras can deal well with glare and other problems that would otherwise require high dynamic range per frame.

Humans would drive better if they could shoot fricken laser beams out of their eyes.

Why would Elon want to model humans when humans are mostly responsible for the problems we want to replace humans for.