A new public data set for self-driving cars shows that even a straightforward city commute may push automated driving to the limit. It also highlights how sharing data could help self-driving cars hit the roads far sooner.
Researchers from Oxford University released the detailed data set, which highlights some of the most challenging issues that self-driving cars will face. The data consists of thousands of hours of data from the same 10-kilometer stretch of road over the course of a year. And it shows how weather, lighting, and even the features of roads themselves can vary wildly in a relatively short period of time.
The researchers tracked the sort of variation that self-driving cars will need to cope with day to day—moving vehicles, cars parked in different ways, and variations in lighting. “Then there are longer-term changes,” says Will Maddern, a senior researcher in the Mobile Robotics Group at Oxford University. “Construction, roadworks, seasonal changes in vegetation, etc.”
When Google researchers began testing autonomous cars, they found that the vehicles were easily flummoxed by busy rotaries. The vehicles, programmed to err on the side of caution, went around for minutes before figuring out how to escape. The Oxford team discovered another kind of problem: during the course of a year, one rotary was moved three times by the city.
Systems that rely on precise mapping, which includes Google’s vehicles, will struggle with such changes. “It’s very much an open problem,” says Maddern. “One of the reasons we collected this data was to find where the systems we are building would break.”
Some companies, such as Tesla, aren’t using detailed maps—instead they’re relying on advances in image and sensor processing to detect and avoid obstacles. But these systems would also be confused by the sorts of features identified by the Oxford team. Google and Tesla lead the way in terms of the amount of driving data collected, but they won’t have as much data as Oxford does showing variation on a single route.
“This data set is a wonderful contribution to the field,” says John Leonard, a professor at MIT who helped develop some of the key algorithms for self-driving cars, and who is working on a research effort coördinated by Toyota. “Large-scale and long-duration data sets can provide a huge boost to the rate of progress.”
Leonard adds that if the companies developing self-driving cars shared their data, it could hasten the arrival of the lifesaving technology. “More generally, I think it would be great if more groups working on self-driving cars could share data sets, and also to make more of the tools available as open source,” he says.
Others in the industry echo his sentiment. Speaking at a conference last week, Gill Pratt, the CEO of the Toyota Research Institute in California, said that given the safety implications, car companies might also consider working together, which might including sharing some of the data they are collecting.
“It’s important to remember that we don’t always have to work alone,” Pratt said. “Our great hope is for constructive competition and also collaboration between all the car manufacturers, IT companies, different governments, and also hardware manufacturers.”