Focusing on tricky situations could accelerate testing of self-driving cars

Self-driving cars are already on the road in various trials, but just when they arrive in showrooms depends on how soon manufacturers can prove their vehicles' ability to handle tricky presented on the road. Currently that involves racking up lots of test miles, which takes time, but a team from the University of Michigan (UM) has developed a new approach it believes can significantly accelerate the testing self-driving cars.

According to the (UM) research team, tests will need to prove with 80 percent confidence that autonomous vehicles are 90 percent safer than human drivers for the technology to be accepted by the public. But to achieve that level of confidence would require test vehicles to be driven in simulated or real-world settings for 11 billion miles (17.7 billion km). Considering it would take over 5,000 years of round-the-clock testing in typical urban conditions to reach that distance, the team set about finding a way to speed things up.

The accelerated testing program the team developed involves finding tricky everyday situations and using them to replace the "uneventful miles" that vehicles experience for the bulk of the their time on the road. In essence, this is designed to put self-driving vehicles through a crash course in handling the most challenging situations they're likely to face in the real world.

The researchers say the huge amount of development miles being covered by the likes of Waymo are currently necessary because tricky situations – from being cut off, to actually being involved in a collision – are relatively rare. Instead of logging bulk miles to experience these situations, the simulation would allow them to repeatedly experience "meaningful interactions" with human drivers. The simulation assumes human drivers are the major threat to self-driving cars, and slots them in at random.

Having run the simulation, the researchers take the results and use them to assess what the likely outcome is when a self-driving car is faced with a certain situation. The results can also be further extrapolated using statistical analysis to predict how the car would perform during everyday conditions.

By condensing testing or simulations down to tough scenarios, the UM team suggests 1,000 mi (1,609 km) of test time could be made as useful as between 300,000 and 100 million mi (483,802 km to 160,934,400 km) on the road. That figure is, of course, speculative – and initial research was conducted using two relatively simple scenarios: an autonomous vehicle following a human driver, and a human driver merging in front of an autonomous car.

Further research is necessary to refine the technique and find out if the accelerated testing process is just as effective at learning about more complex situations like jaywalkers, snow or a car accident up ahead.

The team's white paper detailing the new approach can be viewed here (PDF)

Source: University of Michigan