How to integrate autonomous CPS

Screenshot: United Artists

Dull, dirty and dangerous — those used to be the jobs relegated to autonomous systems. But a decade-plus of improvement in sensor and computing technology has brought autonomy into the mainstream as a defining technology of the future.

At September’s ION GNSS+ conference, I attended a panel titled “Autonomous Cyber-Physical Systems — The Way Ahead.” I came away astounded by how much is changing, and how fast, because of autonomous CPS.

The panel was chaired by John Raquet of the Air Force Institute of Technology and Zak Kassas of the University of California Riverside. It featured presentations covering topics such as the Columbus Smart City Challenge (Dorota Grejner-Brzezinska, The Ohio State University), benefits of precision agriculture (Steve Rounds, John Deere), robotic teammates on the battlefield (Brett Piekarski, U.S. Army), and UAV design and certification (Demoz Gebre-Egziabher, University of Minnesota).

Autonomous cyber-physical systems (CPS) include unmanned aerial vehicles, self-driving cars and unmanned underwater vehicles. The panel addressed the state of autonomous CPS as well as challenges that need to be addressed as we integrate these systems into our environment.

Rather than discuss a specific application, Michael Veth, CEO of Veth Research Associates, tackled a difficult question: Just how much autonomy do we give machines?

“Rigorous risk assessment is the most critical component of machine-controlled autonomous systems,” Veth said. He said the scope of the machine’s autonomous decisions should be limited to the minimum necessary — in other words, avoid the scenario depicted in the movie WarGames.

Another rule: “Don’t put beta software on the street,” he said, recalling the Tesla autosteer system that resulted in a death. Instead, follow DARPA’s example, with its extensive sandbox testing.

Summing up his presentation, Veth provided five guidelines for developing autonomous machines:

1. Perform rigorous risk assessments;
2. limit range of action to the minimum required;
3. use generative models whenever possible;
4. train and evaluate using maximum available data; and
5. always prefer the simplest models.