Skip to main content

A driverless car has crashed: how robots are learning to understand humans

Posted 
A row of Google self-driving Lexus's
A row of Google self-driving Lexus's.()

It's a minor crash on a Californian highway, but a potentially giant moment for mankind: for the first time ever a self-driving car may be to blame for hitting another vehicle. Up to now the humans have been to blame.

The crash shows why it's so hard to develop a self-driving car.

Google says the crash took place on February 14 when a self-driving Lexus tried to manoeuvre around some sandbags in a wide lane.

Reportedly, the autonomous vehicle was traveling at less than three kilometres per hour, while the bus was moving at about 25km's per hour.

The vehicle thought the bus would slow down. But it didn't understand what a human driver knows: bus drivers aren't making way for you.

Three seconds later, as the Google car re-entered the center of the lane, it struck the side of the bus. Thankfully, no one was injured in the car or on the bus.

Google said in a statement on Monday the self-driving car was partly to blame:

We clearly bear some responsibility, because if our car hadn't moved, there wouldn’t have been a collision.

"That said, our test driver believed the bus was going to slow or stop to allow us to merge into the traffic, and that there would be sufficient space to do that."

This article contains content that is not available.

Caption: How self-driving cars navigate. This is not a demonstration of the crash.

Why is it so hard to develop a self-driving car?

The world's first car simulator network for civil applications is a joint project of University of New South Wales and University of Sydney. Here researchers are figuring out how self-driving cars interact with human road users, and how human road users interact with self-driving cars.

Five car simulators worth $40,000 each are housed in a tiny white room in a basement on UNSW's Randwick campus. There's five more at University of Sydney, one in Brisbane, and one in Perth.

Dr Xhitao Xiong and Dr Vinayak Dixit
Dr Xhitao Xiong and Dr Vinayak Dixit in front of a car simulator at the University of New South Wales.()
UNSW car simulator
A car simulator at the UNSW.()

"From a technical point of view I think the tech is already ready," says Dr Xiong, who is in charge of the UNSW simulator lab.

"Several years ago people have already developed enough algorithms, enough sensors."

It turns out developing the algorithms for self-driving cars to move down a highway without veering off the road and crashing is the easy part.

What's much harder is teaching the cars to drive on the same highway as people.

This is partly because people learn to treat self-driving cars differently to other road users. Humans adapt.

Early research suggests the cautious manner and superior awareness of self-driving cars may encourage humans to drive more recklessly, knowing the self-driving car will slow to let them into a lane of busy traffic, or detect them as they run a traffic light.

Basically, among other things, the UNSW and Sydney University researchers have found self-driving cars may make some humans drive more dangerously.

Their work is about introducing the self-driving cars, which follow the road rules and share the same algorithm, to the range of flawed and idiosyncratic human driving.

"To me autonomous vehicles are not vehicles," says Dr Xiong.

They are virtual humans in vehicles."

This article contains content that is not available.

Would you let go of the wheel?

Then there's the hundreds of tiny cues we take into account when driving - things not written in the road rules, but which we all know and follow anyway.

The recent Google car crash is a great example.

Google said it has reviewed this incident "and thousands of variations on it in our simulator in detail and made refinements to our software. From now on, our cars will more deeply understand that buses (and other large vehicles) are less likely to yield to us than other types of vehicles, and we hope to handle situations like this more gracefully in the future."

It's saying that the Google car's algorithm didn't understand buses or trucks are less likely to yield space - something that most drivers know from experience.

Things get even more complicated when a human is inside a self-driving car and has the option of taking manual control at any time.

Professor Mike Regan, an applied experimental psychologist with the ARRB Group, an independent road research agency that helped set up the first self-driving car trials in Australia, says research shows we all respond differently in a crash situation.

"A recent study was done in Canada," he told Hack. "They made the power steering fail. There were 20 different drivers and 20 different responses. Some people let go of steering wheel and the car drove off the road. Other people cursed. Other people acted as if nothing had happened."

You just don't know how people respond."

He said some drivers are wary of relinquishing control, even when they believe the self-driving car is a better driver. "If you give people manual control should you let them take back manual control if the car senses that might not be the safe thing to do?"

Lack of practice and attention may also make us worse drivers. Passive fatigue is the sleepy feeling when you're under stimulated and bored, and it's a big problem right now in the aviation industry, where the average pilot is flying for only three minutes per flight and as a result tend to perform less well in emergencies.

"The biggest worry in the aviation industry is the pilot taking back manual control - they think automated system won't do as good a job as they do.

When the system fails and we have to take control, we may react too slowly, or make the wrong decision.

Even if self-driving cars are safer, Regan expects many people will choose to drive, rather than engaging the auto-pilot.

They might find that incredibly boring."

This article contains content that is not available.

Self-driving cars on trial in Australia

This is important because in three years time you could be inside a self-driving car.

If not that, the car next to you could be self-driving, meaning it could either be autonomous (able to steer, accelerate, and brake, but probably only on highways. A bit like smart cruise control) or entirely driverless (ones that do everything, including parking).

Dr Venayak Dixit, a senior lecturer at the school of civil and environmental engineering, says there is still a gap between the fantasy of self-driving cars, and their reality.

"It's fascinating to hear what the public feel a driverless vehicle is going to be."

But he's also pretty sure the technology is nearing maturity.

It's like a sci fi movie," he says. "It's here."

The first on-road self-driving car trials in the southern hemisphere happened in South Australia in November, after the state passed legislation allowing this to happen. Government hopes the new industry, tipped to be worth $95 billion globally within 15 years, will help the state's struggling manufacturing sector.

The ARRB Group has set up the Australian Driverless Vehicle Initiative, and partnered with Volvo, Bosch and Telstra. Volvo wants to be selling self-driving cars - at least in some countries - by 2017.

There are no Google self-driving cars in Australia yet.

Posted 
Australia, Science and Technology, Automotive Industry