I’m having a hard time believing that this incoherent article was written by a human. If it was, I suspect that it is a job that computers are ready to take over.

Anyway, turning from the dismal text and unnecessarily bland video, there is something of great interest to me here. My own personal autonomous car crusade has taken the form of doing what I can to advance the technology needed for autonomous agents to beat humans at racing.

When I talk about this topic, there are often objections.

  1. People are already convinced. I’m told, mostly by tech people, that no one needs convincing, everyone’s clear that one day robots will drive better than people. Some believe that day is already here.

  2. It’s hopeless. I’m told that people who stubbornly believe that they will always drive better than the best AI will always believe that no matter what evidence to the contrary exists.

  3. It’s science fiction. Autonomous cars will never amount to anything, racing or on the street.

  4. It’s too easy. I’m told that racing, with its controlled conditions and set tracks, is so easy compared to street driving that it is not even worth pursuing.

  5. It’s orthogonal. I’m told that a brilliant racing AI has nothing in common with the tasks of street driving. To these critics it is no more helpful than nuclear reactor design. There is the milder version of this that believes that slow, controlled safety-focused street driving is too fundamentally different from boundary pushing action racing and could be contaminated by its reckless tendencies.

Despite its severe shortcomings, this article and what it is covering corroborate the idea that my focus on racing as a way to improve and hasten the positive reception of autonomous vehicles is a valid and constructive approach.

Many of these points are well refuted. I would add my own logic that refutes both 4 and 5. I believe that racing is an ideal way to test modelling. The idea is that there are many things about driving that you probably should teach a computer to handle outside of public roads. An obvious environment for this would be a virtual one. But how well do lessons learned in a virtual environment transfer to a real one? This is a fundamental and critical question if we are to quickly iterate (to improve) the designs of intelligent agents. By exploring this topic in the limited context of motorsports we can get a sense of how well we can expect improvements realized in virtual environments to transfer to scenarios on real streets.