Friday, February 13, 2015

The four main roadblocks holding up self-driving cars

Read full article
Continue reading page |1|2

SOME day soon, driverless podcars will cluster around our cities, waiting to pick us up on demand. There will be no steering wheel, no brake pedal; once seated, you can take a nap or watch a movie. This public facility will reduce traffic and carbon emissions. Not having to own a car will make transport cheaper for everyone.

Stop us if you've heard this one before.

Why are self-driving cars taking so long to show? For starters, essential technological and social changes needed to make them work might still be decades away. But they are on their way, thanks to some of the world's largest companies. Google has been fine-tuning its autonomous cars for years, amassing hundreds of thousands of kilometres of test drives on Nevada's roads. Last week the developers of the taxi app Uber announced a collaboration with Carnegie Mellon University's Robotics Institute to develop technology for a self-driving taxi fleet.

"It's a very big deal," says Nidhi Kalra, an analyst at the RAND Corporation. "Nearly every auto-maker is pursuing this technology."

So what are the remaining obstacles, and how close are we to overcoming them?

PROBLEM 1

COPING WITH HUMANS

Autonomous cars will confront the same problem that faces all robots designed to operate around people: social interactions are a key part of negotiating our world.

Self-driving cars must be able to navigate situations where drivers rely on eye contact: merging into a busy stream of cars, following the directions of traffic police, allowing another car to pull out in front of them, coping with crowds of people on the street.

They must also be able to interact with their passengers. Semi-autonomous cars must detect if their drivers are nodding off during long stretches of autonomous driving, and then be able to rouse them back into paying attention.

This transition between human and software control, known as the handoff problem, is a big challenge in automating all kinds of processes. The crash of Air France flight 447 in the Atlantic Ocean in 2009 was the result of the crew responding incorrectly when the autopilot suddenly handed control to them.

HOW CLOSE ARE WE?

Human-robot interaction is hard, and our ability to do tests in this area is limited by the number of robots already out there performing real-world tasks.

Giving cars the kind of social intelligence displayed by Jibo, a robot designed for domestic use (New Scientist, 19 July 2014, p 21) could help them navigate our world safely. The other option is to make the system completely autonomous. This already happens in self-driving trains the world over, the first example of which was launched in Japan in 1981. Trains are much easier to automate because they are constrained to travel on tracks, and there's no need to interact with other train drivers. To be truly useful though, self-driving cars ought to be just as woven into our lives as existing vehicles, maybe more so. How that will work is still up for debate.

DIFFICULTY RATING

* * * * *

PROBLEM 2

THE WEATHER

California's sunny highways are great for Google's self-driving car. But how would it have fared in the snowstorms that hit the US east coast earlier this month, or in protracted drizzle in London?

Badly, is the answer. A self-driving car competition hosted by Hyundai in South Korea last year provides a perfect example. On the first day of the competition, the cars handled the track and obstacles at speed with few mistakes in dry, clear conditions.

The following day was wet and misty, and the same cars missed turns on the course, veered into the opposing lane and mounted the kerb. Machine vision systems that use ordinary cameras can't see as well in these conditions, and this trips up the cars' navigation software.

HOW CLOSE ARE WE?

Equipping cars with a wider array of sensors, as well as embedding sensors into the environment, will help. In particular, cars can be fitted with devices to tap into regions of the spectrum that cut better through fog or rain. "I don't see weather as the showstopper that people are talking about," says Kalra.

Testing the cars in more challenging climates than Silicon Valley's will help, too. To that end, the University of Michigan is building a 13-hectare test site to put cars through their paces in the rain, sleet and snow of the Midwest. Communication between cars and roadside sensors will be one of the lab's main focuses, as weather conditions can't throw this out so easily.

DIFFICULTY RATING

* * *

PROBLEM 3

SECURITY

It's one thing for a virus to cause your phone or laptop to crash, but if the same fate befalls a driverless car, the consequences will be much more dramatic.

Any software bugs in an autonomous car might let a hacker take control of the vehicle remotely, perhaps locking the passengers inside until a ransom is paid. Car companies will probably pore over their software more intently than the average app developer, but they will still miss things. When problems crop up, they will need to fix them as fast as possible. It should also require minimum effort by the user, as the hassle of dealer visits means passengers may settle for cars with unpatched software.

Read full article
Continue reading page |1|2

If you would like to reuse any content from New Scientist, either in print or online, please contact the syndication department first for permission. New Scientist does not own rights to photos, but there are a variety of licensing options available for use of articles and graphics we own the copyright to.

post from sitemap

No comments:

Post a Comment