Fully self-driving cars will need to handle lots of variables: Harassment-prone humans, temporary detours, and animals in the road, to name just a few.
And while autonomous technology currently can deal with aggressive drivers and construction zones, it may not be able to avoid killing critters if they want to keep those inside a self-driving car safe from harm.
According to the Insurance Institute for Highway Safety, about 200 people are killed each year in vehicle collisions with wildlife. State Farm reports that drivers in the US collided with 1.25 million deer alone in 2014, causing $4 billion in damage to vehicles—and a mess on roads. The Virginia Department of Transportation spends $4 million per year disposing of roadkill.
While millions more animals die every year in collisions with vehicles, most of these are non-issues for drivers since their cars typically run over small creatures like rodents and smash into birds. But the autonomous technology company nuTonomy recently found that even flocks of seagulls can cause problems for self-driving cars.
As nuTonomy's cars have logged miles gathering data while driving around the Raymond L. Flynn Marine Industrial Park in South Boston, its tech has had trouble recognizing flocks of seagulls that gather on the roads. "One bird is often small enough that the car assumes it can be ignored," nuTonomy CEO Karl Iagnemma, told the Boston Globe.
"But when you have a flock of birds together, it looks like a big object, so we've had to train the car to recognize the birds," Iagnemma added. "It wasn't something really that we've seen before."
NuTonomy's autonomous tech reacts to flocks of seagulls the way a human would: By slowing down while approaching the birds until they scatter and fly away. Iagnemma said that the company is teaching the car's software to recognize the birds and "give the vehicle a better capability to predict what's going to happen next" by feeding imagery of flocks of gulls into an algorithm.
But as anyone who has ever encountered an animal on the road knows, figuring out a squirrel, deer, or flock of bird's next move is the hard part.
Nature Is Imperfect and Unpredictable
As Smithsonian.com noted in 2015, "the main challenge is that nature is imperfect and unpredictable, and it isn't clear yet how the rigid calculations of computers will handle the sometimes erratic behavior of animals."
"Even if we develop the perfect automated recognition and avoidance system, you still have an imperfect ecology and wildlife behavior system," Fraser Shilling, the director of the California Roadkill Observation System, which collects data on roadkill and pinpoints collision hotspots to reduce wildlife carnage on roadways, told Smithsonian.com.
Even though Volvo developed a Large Animal Detection with Autobrake system it introduced on the 2017 S90 sedan, the technology still can't predict a moose's every move. "We can only make a rough prediction of the motion of the [animal] based on its current position and speed," Erik Coelingh, Volvo's senior technical leader, told Smithsonian.com.
Related
- Who Should Your Self-Driving Car Save in a Crash? You or Pedestrians?Who Should Your Self-Driving Car Save in a Crash? You or Pedestrians?
Andy Alden, a researcher with the Virginia Tech Transportation Institute, told Smithsonian.com that a study his organization conducted with Toyota found it would be difficult to for autonomous tech to predict animals' actions. But he added that "there certainly are some things you could drop into an algorithm, like time of day, time of year, the kind of environment along the road, the width of the road, the amount of traffic on it" to help avoid animal collisions.
That's exactly the kind of data that nuTonomy hopes to gather in South Boston and by confronting flocks of seagulls. "You end up having to adapt the system to the local conditions," Iagnemma said.
"It's a good example of a unique aspect of driving in Boston," he added. "Is it a fundamental change to our approach? Absolutely not. Is it something that takes a little time to adjust our software to?