Director of Mcity, Dr. Huei Peng, Presenting a Talk on “Physics Enhanced AI”
Perception, path planning and control are three important tasks in developing automated vehicles. For level-2 automated vehicle functions such as super-cruise (GM’s trademark) or Autopilot (Tesla’s trademark), the fundamental functions include lane/curvature detection, obstacle detection, path planning, and steering/speed control. Existing methods in the literature can be largely divided into two categories: end-to-end, and step-by-step. End-to-end approaches compute control actions directly, while in the step-by-step approaches, the tasks are separately conceived, designed, and validated. The step-by-step approach dominates the automotive field for many decades, until the new kid on the block, the artificial intelligence (AI), emerges, promising to totally revolutionize all aspects of the modern society, including design autonomous vehicles. AI concepts were promoted by “tech” companies such as Nvidia. Many of the end-to-end promoters believe it is not necessary to use vehicle dynamics/physics based knowledge and models—the final control decision can be figured out by the “intelligence” itself. A fundamental question this proposal aims to answer: is it true that dynamics/physics are yesterday’s knowledge and is becoming obsolete by the AI approach? Or, they still have some value in the design of autonomous vehicle control systems? If so, how do we use these two approaches in a complementary way? Is there a systematic way to identify tasks more suitable for AI vs. those more suitable to physics/dynamics?
Ann Arbor SPARK Central Innovation Center
East Liberty Street
Ann Arbor, MI, USA