Tesla is holding Autonomy Investor Day.
The first part of the Autonomy day described the new hardware. This second part is the neural net vision system description and the interaction with fleet data. The third part self-driving software development.
Tesla has a fleet learning approach. They can ask the fleet to provide examples of different examples of specific situations. They can get hundreds of thousands of examples of cars that are carrying bikes or road debris.
They get feedback from the fleet of human interventions that were needed or times when the system was confused.
They are able to accurately path predict what will happen ahead on the road and the expected movement of drivers, cars and objects.
LIDAR points have less information than vision. LIDAR is a shortcut to self-driving and sidesteps key issues that are needed with vision-based driving. Powerful vision recognition and proper understanding and processing of the information is the path that Tesla is taking.
Tesla has a sophisticated trigger infrastructure to select the right cases that the fleet is experiencing that require improvement.