Until recently Tesla was using the Mobileye vision system for their Autopilot software. This system has several features, such as sign recognition, obstruction/vehicle/pedestrian recognition and lane tracking. The one at issue here is the lane tracking.
Lane tracking is a very treacherous capability. It's quite easy to develop a vision system that appears to work. It's easy with freshly painted lane markers and no exits or merges. Complexity and special cases quickly pile up with right hand exits and entrance merges. Quick fixes such as always following the left-side lane line come back to bite you when encountering left exits and lane merges. Roadways with faded and old ground-away lane markings that are confusing to human drivers are even worse for lane-following vision systems.
Tesla mitigated this by requiring each section of highway to be successfully observed several times before enabling Autopilot. I suspect that here we are seeing the weakness of that approach -- the roadway marking degraded below a threshold and the lane tracking made a bad decision. This has probably happened thousands of times before, with the harmless result of taking the wrong exit or driving in a break-down lane. Here it was a fatally bad result of driving into a barrier centered in the 'false lane'.
Biting the hand that feeds IT
