It can’t, nor can most cars with automatic braking. The Tesla will happily run headfirst into a brick wall all day long. The reason is that the driving world is chock-full of objects with zero velocity relative to terrain - the trivial cases being a rise in the road ahead, or an at-grade bridge where the travel lanes suck below grade to pass. Therefore, autopilot and many auto-braking algorithms filter out completely static objects (see previous story about a Tesla ramming a fire truck stopped on the highway).
Musk has commented publicly before (though somewhat obliquely) about this flaw. He indicated that the company is trying to build a map of reference data so that it can be filtered out automatically, and real hazards can be seen/found.
Mapping can never solve this problem. If these cars don’t yet have the ability to detect stationary physical barriers that represent a crash risk, then they are further away from being practical than I thought, and I’m extremely pessimistic. It is clearly a hard problem to solve with naive tech. Unfortunately we have a huge industry and tens of billions of dollars and politicians, corporations, governments, and media who all believe this naive tech is close to perfect. But it can’t see a wall or a fire truck in its path? If this isn’t solved then this whole house of cards will fall apart. The sooner the better, IMO. Let’s start building cars that supplement driver awareness instead of numbing it.
This would be trivial to solve with a vertical mounted LIDAR. You vertically mount it and if the horzontal distance measured level with the bumper of the car are significantly closer than the horizontal distance at a lower angle you can classify that as a stationary object, if it's a continuous lengthening or shortening then it's a ramp. The classification is almost trivial (though the hardware may be expensive or difficult to implement for high speeds).
Musk has commented publicly before (though somewhat obliquely) about this flaw. He indicated that the company is trying to build a map of reference data so that it can be filtered out automatically, and real hazards can be seen/found.