| So there's a couple things at play. Lidar range is more than double visual range, in practice. When figuring out collision avoidance paths for an object you actually end up approximating some np hard problems to find a path that won't have collisions and won't be too "careful". This ends up being fairly computationally intensive, and adding the extra time significantly improves your planning. Doubling compute time tends to beat doubling your training dataset in terms of system quality, at this scale. Extra time also turns a number of situations from "guaranteed kill" to "we can avoid the accident", because the car is traveling really fast and those extra seconds can be used to brake, find a new path, etc. In visibility impaired situations, lidar and vision have different constraints and ways they fail, and the intersection of the two can significantly improve scene understanding ( see waymo's snow demo ). In a lot of cases, path planning can be dramatically improved by having maps. If you're going into a curve and know the shape of the road, you can preload that and spend your time on more important tasks like object detection and path planning. Etc etc etc. This is absolutely not a domain for intuition and thought experiments. The pragmatics of the industry are highly intricate and responsive to constraints that are only visible if you've worked on this stuff. |