The incident of Waymo’s automated driving taxi happened few days ago. It reminds me about my personal opinion about automated driving technology obstacles that I’d like to recall here.
The machine learning (ML)-based software modules, by their nature, such as object detection and tracking (using camera or Lidar data inputs) are not reliable and robust. When we train a model, achieving 90% of accuracy is considered very high. It is great for many applications. However, it is nothing for safety-critical applications like self-driving cars. For example, the probability of making a wrong detection within 1 hour of operation is 1. Please have a look at this post for details, and my proposal for asserting ML-based software module.
The high definition (HD) maps, there is no ways to update the HD maps in real-time w.r.t the changes of the roads, building, landmarks,… We see in this case of Waymo’s automated driving taxi, it seems that the planning algorithms rely on HD maps for making decisions. And the road cones and the HD maps do not agree each other, thus, it makes the car confused. Thus, HD maps should be complementary, not main source of making decisions.