The reason why Tesla FSD performs better with just computer vision than systems that have LIDAR and radar is a bit counter intuitive. Allow me to try to explain. In the early days of self-driving, people thought if you built a really accurate model of what's happening around the car, then you can come up with algorithms to determine the best way to control the car through the environment. On the surface, it seems like LIDAR and Radar would allow you to drive better as you can more precisely measure speed, distance, etc. In reality, with deep learning computer vision has gotten pretty damn accurate at determining speed and distance — so much so that the added benefit of extra sensors is negligible. When you're changing lanes, it doesn't matter if that car that's approaching you is 100 feet away or 95 feet away. What matters is, do I have time to change lanes or not? As anyone who has tried the latest FSD can see, computer vision has long since been accurate enough to make these determinations. But deep learning gives you another super power that turns out to be critical for self-driving. While a great perception system can tell you what's happening in the present, what the self-driving car actually needs to do is predict the future. Not just how far away is that car, but what do you think it's going to do next? Is that pedestrian at the crosswalk going to cross, or are they going to stand there and wait? Subtle things like their facial expression, direction of gaze, etc that can be predictive are present in the visual spectrum and not with other sensor modalities. So deep learning and predicting the future, not just perceiving the present, are the key to safe driving. Humans don't think about this because if we see what's in front of us we can immediately and intuitively predict what pedestrians, cyclists, and cars will do. But that's not so for computers. In order for deep learning to work, you need a big, diverse dataset and a ton of compute. It turns out it's a lot easier to build this on camera data than with sensor modalities like radar and LIDAR. Because cameras are so low cost, you can build them in to large fleets of millions of cars and build massive datasets. The cost and technical differences between other sensor modalities make it much more difficult and costly to build a large scale dataset. The training process becomes more complex too. Although camera images seem inferior, they work a lot better with deep learning. And deep learning is what unlocks cutting edge AI. That is why, counter intuitively, these pure vision systems are resulting is smoother, more comfortable, and more humanlike drives than the traditional more spatial approach to self-driving. Every human you see driving around is just using data in the visual spectrum. They don't have radars and LIDARs measuring the exact speed of every car. They just look around and know if it's safe to execute a maneuver or not. I know it seems hard to believe that a simpler and cheaper set of sensors could actually perform better, but that is objectively what has happened. Go take a ride for the latest FSD if you don't believe me and see it for yourself. Compare to Waymo and you will see what I mean. The way I see it, computer vision is necessary and sufficient for self-driving. It is necessary, meaning you can't do self-driving without computer vision. If your computer vision system fails but you still have LIDAR and radar active, you won't be able to read traffic lights, signs, or lane lines — you can't drive. So it is necessary for self-driving. It is also sufficient. Computer vision alone, without any other sensor modalities, is enough to drive a car safer than a human. We know it is sufficient because all humans drive using just vision. But Whole Mars, you ask, wouldn't adding other sensor modalities make it even safer? Of course, in some cases additional sensor modalities may be helpful. But you need to design the vision model at the core of the system to be able to drive the car safely if all else fails. Let's say a big rock falls off a cliff and knocks out your LIDAR. The car needs to still be able to drive safely to the service center even without it. Let's say the radar accidentally gets unplugged somehow, or is getting some interference — the car still needs to be able to drive even if these additional sensor modalities fail. So in any case, you want a vision system at the core that is strong enough to handle the entire driving task safely on its own if all else fails. Once you have that, you can layer on additional sensor modalities on higher end vehicles if there's a benefit to doing so — but if those sensors fail, vision still needs to be strong enough to drive the car on its own. Hope that makes sense. Let me know if you have any questions.