... the limited videos about the Uber death, the Uber SDC software was by no means ready for driverless control, or even testing on city streets. Its inability to react to the pedestrian in front of it should have been evident if it had been tested properly, which it clearly wasn't...
Not to defend Uber, but they would have already tested this very basic function. That is, their lidar had to be able to detect pedestrians or a vehicle in front, or in a collision path with their car. They would not have been able to have so many test cars out on the public streets for that long without this basic ability. This Velodyne lidar is the same one used by many SDC developers. So, sensor performance is not the problem.
Something went very wrong on that unfortunate night. Perhaps the sensor failed and the software did not know that it failed. Perhaps the software had a glitch. For example, multi-threaded software may have a thread stalled, and did not perform all processing on all the information that it received. That would be similar to a person doing some cooking, and while receiving a phone call forgot about the pot being left on the burner.
For a system to get past the experimental phase, it has to be able to self-detect a sensor or subsystem failure, then to take graceful degradation action, such as pulling over and slowing down to a stop. Of course, a critical sensor or subsystem may fail in such a way that the system cannot even make a safe stop. Think processor failures or power supply failures.
Hence, they will need redundancy for all critical subsystems. The testing of the ability of the system to handle these failures can be done most economically in the lab, and does not have to be done in the streets.
The developers only go out in the streets to test their sensor performance in real-world situation, or to see if their system brain can handle the more complex scenarios, such as being surrounded by other cars, bicyclists, pedestrians, etc... This is the kind of stimulus that is very difficult to simulate in the lab. And I do believe it can be safely tested in the streets because of the slow speed to allow for human overriding. Waymo has been doing it for a few years, using a toy car that could go only 25 mph.
So, something fell through the crack. A failure in hardware or software occurred, and they did not detect it, or knew how to handle it.
The failure of the Teslas has been mostly sensor performance. For example, look at the video I linked earlier of an owner testing the ability of his car to take an exit ramp (post #1161). He correctly observed that the left lane marking was covered up by gravel at 1:10, and the car needed that marking to follow the road. The vision system was not smart enough to tell where the road was without the lane markings, while humans have no problem.
So, don't discount the ability of the humans to make sense out of things that the computer cannot, despite all the hype about AI.
And that is just to keep the car on the road, and nothing yet about watching for other cars, pedestrians, bicyclists, stray objects on the road, reading street signs, etc...
