What happens if something goes wrong, if there’s software or component failure for a Level 4 automated truck?
It’s one of the first questions that the Daimler Trucks team approaches, as it’s paramount to safety. Data plays a huge role in testing a neural net’s operational safety.
“Torc is of the opinion the quality of a data set is more important than the quantity of the data set,” Michael Fleming of Torc Robotics said. “Now the challenge is to ask, what is a quality data set? If you collect quality data in one region, is it quality data in another region?”
“While neural nets can do some amazing things,” he continues, “they also fail quite a bit with some of the simpler problems that you and I would consider to be obvious. And one of the reasons behind that is we—and not just Torc, but the entire industry—don’t exactly know what’s going on within highly complex neural nets. Now the challenge is how do you know if your neural net is safe if there are an infinite number of scenarios to test against.”
Torc has already won one of those challenges, and the company did it 12 years ago during the DARPA Urban Challenge. That’s not to say they won the race (they came in third) but their automated driving victory was far more important than the ribbon color:
“In the DARPA Urban Challenge, the vehicle had a handful of sensor failures during the actual race,” Michael said, noting that the requirements of the race stipulated that a human driver was not allowed in the vehicle, and there was no remote control. “Our vehicle detected those failures, turned on its blinker, moved to the side of the road, and then the computer went through a diagnostic routine to classify and correct that error. Our vehicle received a clean bill of health through our health monitoring subsystem. Then it activated its left turn signal, re-entered the road, and continued the race.
“We’ve taken that same mindsets and rolled it out in our other customer applications over the last 12 years. Failures will take place in any system, in any technology, that you create, but what’s important is the ability to detect that failure and to be able to fail safe.”