If I understand it correctly, auto-pilot systems like that are based on recognizing and reacting to road and traffic conditions. If the car doesn't recognize the road conditions and the driver doesn't take over, disaster can happen - such as happened several years ago in a deadly crash in Mountain View. In the Houston crash the auto-pilot system probably did not recognize the sharpness of the curve (or that there was a curve). When lithium batteries catch fire it gets nasty.
The car itself (Tesla Model 3) has won awards for safety - consistently beating ICE vehicles. Just like a car with a ICE, if it catches fire, it's going to burn.
The AutoPilot feature uses front-facing radar and a number of cameras to keep the car centered in the lane and to avoid running into the car in front if it slows down. It can also recognize a number of objects like traffic cones and barrels, signage on the side of the road and painted on the road, traffic lights, etc. But the system is designed to be used always with a driver monitoring and ready to assume control at any point. It's just like cruise control in that you can't just turn it on and ignore the road.
As far as avoiding obstacles and keeping the car in the lane, it's really fantastic. It does like to take curves faster than I personally would take them, and that makes me nervous sometimes, but I am a very cautious driver. It's great for lane changes on the highway - I just hit my blinker, and I look too, but it's an extra level of safety having the car choose a safe time to do the lane change and then execute it. It's good also at watching traffic merging into the highway and avoiding a conflict. My biggest issue with it is that it infrequently does this "phantom braking" where it thinks there is some reason to slow down and so it does that, sometimes quite assertively, on the road. I've had to tap the accelerator when it did that out of concern that some following car might rear end me.
The Tesla fan-boys disagree with me on this, but I don't think it's anywhere even remotely close to a fully autonomous driving system. I don't see cars being able to do that until they have higher order cognition. They use simulated neural nets that work in a way modeled after the visual cortex in humans. Those are great at pattern recognition, but they have no higher level thought process. Even if that's only needed 0.0001% of the time, it's still needed, and not having it is likely going to make people not comfortable to use the system - even if it has an overall safety rating better than a human.