Autonomous Driving - Not Happening Any Time Soon

There are six levels of driving automation (see Wikipedia for details):

  • Level 0: No automation
  • Level 1: Driver assistance
  • Level 2: Partial automation
  • Level 3: Conditional automation
  • Level 4: High automation
  • Level 5: Full automation

We have Level 2 today but it mostly still sucks but will probably improve. There are Level 3 cars but they operate under very tight constraints (e.g. in traffic jams) and I tend to agree with Don Norman that it’s not a desirable model. What the media is really talking about and what most people tend to associate autonomous driving with is at least Level 4, but actually more like Level 5: You can lean back and watch a movie while your car takes you from A to B. I’m saying we are nowhere near to this happening.

The major factor that caused the current acceleration of research into autonomous driving is the recent progress observed in machine learning systems. While the media has you believe that your job will soon be replaced by an artificial intelligence, most of that progress concerns pattern matching like image recognition and voice recognition as well as pattern generation like style transfer.

Have a look at the architecture of an autonomous driving system (taken from this paper):

Architecture of autonomous driving systems

In this model, machine learning is at best applied to steps 1 and 2. A lot of work is being done improving the detection and tracking of objects and building a 3D model from that information. But deriving a strategic plan for when to turn, when to change lanes, when to slow down and when to accelerate, and so on, is still up to the algorithm developed by the teams at Waymo, Tesla, Uber, et al.

It can’t be machine learning: How would we be able to ensure that our machine learning software wouldn’t violate the most basic rules, e.g. don’t kill humans, don’t drive on the wrong side of the road, or don’t crash into other cars? Adverserial attacks are easy to create and we cannot verify the correctness of a neural network. So even if we put “AI” into the motion planning step, we can only hope that our machine learning model is not going to kill us.