Backseat driving can irritate even the most stoic of human drivers, but for autonomous vehicles those nagging comments could lead to the perfect road technique.
The software that powers autonomous vehicles is often trained on hours of footage of people driving. This footage normally only contains human drivers at their best, but this could mean that if an autonomous vehicle is faced with a pileup ahead, for example, it will only know how to stick in lane rather than swerving out of the way.
So Jacob Beck and his colleagues at聽Brown University in the US have taught autonomous driving software to learn from both good and bad driving examples, with the help of constructive criticism from a backseat driver.
Advertisement
In a simulator, the team recorded a human driving well, swerving, and changing lanes as if to avoid something 鈥� with each task lasting 20 minutes. A backseat driver labelled the footage positively or negatively, to show proportionately how well the driver was performing. In total the team collected 21,000 labelled stills.
The software absorbed all of this information, like a child learning from bickering parents, before taking a spin itself in the simulator.聽The team recorded times until either the car crashed, or the car had driven off the road. It performed over 1.5 times better than when no initial feedback was provided. Here at least, backseat driving had some merit.
鈥淪ometimes seeing bad behaviour and knowing it鈥檚 bad can be very instructive,鈥� says Michael Littman at Brown University, a co-author of the paper. Next, the researchers would like to try teaching the system to cope with a safety driver taking over control鈥攖he kind of backseat driving no right-minded human would stand for.
Reference:
