Meanwhile, the competition who is using LiDAR has FSD cars. You're understating the importance of this sensor.

You can train a DL model to act like a LiDAR based on only camera inputs (the data collection is easy if you already have LiDAR cars driving around). If they could get this to work reliably, I'm sure the competition would do it and ditch the LiDAR, but they don't, so that tells us something.

It is very true and worthwhile to point out that the only company deploying L4 at scale is using LIDAR. And that company is not Tesla

The mental gymnastics Tesla fanboys use to explain this away are incredible.

The Tesla social media team actively used to post positive spin on comment threads. It wouldn’t surprise me if they have LLM doing this now.

Researchers had this knowledge in 2007, when the only cars to finish the DARPA Urban challenge were equipped with Velodyne 3D LIDAR. Elon Musk sent us back a decade by using his platform to ignorantly convince everyone it was possible with just camera alone.

For anyone who understands sensor fusion and the Kalman filter, read this and ask yourself if you trust Elon Musk to direct the sensor strategy on your autonomous vehicle: https://www.threads.com/@mdsnprks/post/DN_FhFikyUE

For anyone wondering, to a sensors engineer the above tweet is like sayin 1 + 1 = 0 -- the truth (and science) is the exact opposite of what he's saying.

Isn’t this article about Tesla admitting their system is as good as it’s going to get? They’re changing their definition of FSD to pretty much “current state”.

It's an extra sensor you'd add into the mix, you'd still have cameras. Like the radar sensors. I think the reason Teslas don't have it, is because the sensor hardware was expensive a few years back. I assume they are much cheaper now.

Tesla have also backed themselves into a corner by declaring that older models are hardware capable of FSD, so they can't easily add LIDAR to new models without being liable for upgrading/refunding previously sold models.

I thought they had it on some models already, then removed it on models after?

edit: no, it was ultrasonic sensors. But this was likely object detection, and now it's gone.

New for 2027 - ABSOLUTE Self Driving Pro Max!

I think you might be under-estimating the importance of not hitting things.

If you look at the statistics on fatal car accidents, 85%+ involve collisions with stationary objects or other road users.

Nobody's suggesting getting rid of machine vision or ML - just that if you've got an ML+vision system that gets in 1 serious accident per 200,000 miles, adding LIDAR could improve that to 1 serious accident per 2,000,000 miles.

Because LIDAR can detect the object at the beginning of the perception pipeline, whereas camera can only detect the object after an expensive and time consuming ML inference process. By the time the camera even knows there's an object (if it does at all) the LIDAR would have had the car hitting its brakes. When you're traveling 60 MPH, milliseconds matter.

Just to put numbers on it, 10ms at 60mph is just under a foot. I don't think that matters too much, but if we're talking 200ms that's 10-15 ft which is substantial. I have no idea how long the ML pipeline is, though.

I don't know about you but one of my primary goals when driving is not hitting into things