Automotive

After just 20 hours of training, Wayve's fast-learning AI car is already driving itself on unfamiliar roads

After just 20 hours of trainin...
Wayve co-founders Alex Kendall and Amar Shah. Their fast-learning AI autonomous car is now doing short trips on unfamiliar roads in the UK.
Wayve co-founders Alex Kendall and Amar Shah. Their fast-learning AI autonomous car is now doing short trips on unfamiliar roads in the UK.
View 2 Images
Wayve co-founders Alex Kendall and Amar Shah. Their fast-learning AI autonomous car is now doing short trips on unfamiliar roads in the UK.
1/2
Wayve co-founders Alex Kendall and Amar Shah. Their fast-learning AI autonomous car is now doing short trips on unfamiliar roads in the UK.
The car uses only cameras and a GPS, and the only human input it gets is "no, don't do that" in the form of corrective maneuvers. Beyond that, the whole thing is based on deep learning.
2/2
The car uses only cameras and a GPS, and the only human input it gets is "no, don't do that" in the form of corrective maneuvers. Beyond that, the whole thing is based on deep learning.

Forget large arrays of sensors and radars. Forget hard-coded road rules. British startup Wayve taught a car to teach itself to drive, and using only some cameras, a sat-nav and 20 hours' worth of experience, it's already driving itself short distances on unfamiliar UK roads.

When these guys first put their control hardware, camera, computers and AI learning software into a Renault Twizy-based StreetDrone back in July last year, the result was a machine that went from literally not knowing what the pedals and steering wheel did, to being able to follow a lane more or less indefinitely, in the space of just 20 minutes.

The system learns not by being told what to do, but by being told when it's made a mistake. each time a human driver has to grab the wheel and steer it back on course, the system considers that a penalty, and tries to learn from what its cameras were seeing just before the event to figure out how to get better next time. On the other hand, it thinks of any time spent driving without a human intervening as an accomplishment, and gives itself a virtual cookie, thus reinforcing good behaviors and working to eliminate bad ones. The team says it "propagates uncertainty throughout the model" to help it learn faster and more efficiently.

And now it's out on the road. The company didn't code the car with any road rules – it simply followed the same "intervention equals penalty" approach. The Wayve car now has its own GPS system to add to the mix – a low-res, consumer grade thing that runs on a smartphone – and it's worked out through trial and error that proceeding along the planned route on the GPS is cookie-worthy.

It also appears to have figured out that it's not supposed to bang into other cars or cyclists, that it needs to give way to other cars at T-intersections, and that green means go when it sees a traffic light. The team has furnished a short video of the car driving a few blocks around Cambridge, England, on unfamiliar wet roads, without needing any intervention between the start and end points.

It's fair to say the Wayve Twizy would be a pretty scary car to ride in, with its jerky steering motions (potentially a result of that uncertainty propagation) and given the fact that it's literally making this all up as it goes along. It would also be a frustrating car to drive behind at this point, since it behaves pretty nervously and never breaks more than 11 mph in the video.

The car uses only cameras and a GPS, and the only human input it gets is "no, don't do that" in the form of corrective maneuvers. Beyond that, the whole thing is based on deep learning.
The car uses only cameras and a GPS, and the only human input it gets is "no, don't do that" in the form of corrective maneuvers. Beyond that, the whole thing is based on deep learning.

But the punchline is that the car has spent only 20 hours learning to drive at this point. And while it's certainly not doing as well as you'd expect a human driver to do after 20 hours of practice, it's worth remembering that humans have other humans telling them what to do instead of figuring everything out from scratch – not to mention more than a decade and a half sitting in passenger seats quietly absorbing road rules, etiquette and a stream of powerful language that can be used to signal the breaking of that etiquette.

Compared to, for example, Tesla's Autopilot, which is far and away the most advanced system currently deployed on the road, the Wayve car is slow, wobbly and untrustworthy. But it's in its absolute infancy as a technology, and will progress quickly as the hours pile up. And one reason to hope it does well is that it runs on hardware that costs less than a tenth of what other autonomous car platforms roll with.

Wayve is confident: "We're going to be the first to deploy autonomous vehicles in 100 countries." But to bring things down to Earth a little, the company has a long, long way to go before it becomes a serious contender. Autonomous cars from Tesla, Waymo and any number of other companies have been driving around, learning and gathering data for millions of miles now, and it's not the regular driving bits that are keeping these tech giants from releasing Level 5 JohnnyCabs.

It's the once in a hundred thousand miles occurrences, the times when the car can't discern lanes on the road, when traffic behaves in new and unforeseen ways, when visibility sucks or the road is slippery or a cat jumps out. When there's weird roadworks conditions, detours, paint spills, potholes, fallen branches, accidents, sunset in your eyes or a thousand other conditions that the Wayve car might simply never encounter in testing – at least, not enough to work out a comprehensive set of ways to deal with things.

Wayve's solution to this, it appears, is simulation. Because all it's using as inputs is a camera and a GPS app, it's easy to fake up scenarios and run the car's brain through endless simulations of any situation the creators can dream up, thus accelerating its learning curve even when it's not on the road.

Certainly, we're impressed with what this company has achieved in less than a year behind the wheel, and we hope Wayve gets together the resources and capital to get this thing to the next level and, eventually, to the mythical Level 5 autonomy we keep hearing is just a few years away. Check out the video below.

Source: Wayve

Urban Driving with End-to-End Deep Learning

7 comments
Brian M
Be fun to see what happens if it does the 20 mile ferry trip across the English Channel to France where they drive on the opposite side of the road! Or maybe not.....
WilliamSager
Based on the latest news stories the real deal breaker may come down to the AI's being racist or not.
FabianLamaestra
Wow, the steering wheel algorithm is very nervous. That would make the passenger very nervous as well.
jd_dunerider
Seems like a good way to teach AI, as long as the teachers have a moral compass and good set of guidlines they follow.
Trylon
"the mythical Level 5 autonomy we keep hearing is just a few years away." Spare me the snark. I've never seen anybody claim that it was just around the corner. What I do know beyond a shadow of a doubt is that it's coming whether you like it or not. With every passing month and year, the software and systems get ever better, which is more than I can say about human drivers, who seem to get worse over time, both individually and collectively. Not a day goes by when there aren't hundreds of news stories about bad and often fatal accidents involving cars, and who knows how many tens of thousands of fender benders and close calls.
Gill Davis
Humans make only 5 decisions a second, so try you commands at a slower rate of hertz for smoother reactions
ljaques
Is a full eleven kph a big milepost? I tucked my elbows in when he went by the car and the bike. Did it even see the bicyclist? It had no reaction. The steering method needs some work in smoothing, and it doesn't seem any better now than on the first dirt road test, IIRC. I'll bet it felt like a nervous Nellie driving for the first time.