Wayve Taught a Car to Drive Autonomously in Just 20 Minutes

Home > News > Content

【Summary】A team of researchers in Britain under the company name of Wayve, showcased its “reinforcement learning” algorithm that, when used in tandem with a human safety driver, taught a vehicle how to drive autonomously in just 20 minutes.

Original Vineeth Joel Patel    Jul 17, 2018 8:30 AM PT
Wayve Taught a Car to Drive Autonomously in Just 20 Minutes

Artificial intelligence is one of the major keys that will take autonomous cars from being Level 2, which are already on the road today, to Level 5, which refers to fully self driving vehicles. Automakers and technology companies have run into a dilemma with artificial intelligence, as experts believe that the adjustment periods that AI systems need to follow to operate on their own are a lot longer than everyone expects.


Automakers are experts at making vehicles and technology companies have started to dabble in artificial intelligence systems that are specifically geared towards automobiles. That's one of the obvious reasons why traditional company's, like Ford, have teamed up with tech giants like Baidu. 

Ford isn't the only company to set its sight on a partnership to ensure that its driverless cars are fitted with the right artificial intelligence systems. Daimler recently chose Silicon-Valley based Xilinx as the company for its AI-based automotive applications. 

Clearly, if you want autonomous vehicles to operate properly, they have to have high-tech AI systems. A new team of researchers out of Britain just revealed why having the right artificial intelligence system is crucial, as they developed a system that can teach a car to drive on its own in just 20 minutes. If that's not impressive in today's modern times, I don't know what is. 

Wayve Autonomous 2.jpg

Wayve Comes Out With New, Innovative Tech

As New Atlas reports, two artificial intelligence Ph.Ds from Cambridge University have gone all-in on machine learning and have founded a new company called Wayve. The company recently released a video of the Renault Twizy, which was modified with the necessary components to drive on its own, teaching itself how to drive over a span of 20 minutes. 

According to Amar Shah and Alex Kendall with Wayve, the industry has focused too much on hand-engineering to find a solution to the autonomous car issue. 

"The missing piece of the self-driving puzzle is intelligent algorithms, not more sensors, rules and maps," said Shah, the co-founder and CEO of Wayve. "Humans have a fascinating ability to perform complex tasks in the real world, because our brains allow us to learn quickly and transfer knowledge across our many experiences. We want to give our vehicles better brains, not more hardware." 

With that philosophy in place, the Wayve team took a Renault Twizy and outfitted it with a couple of components. Those included a camera at the front of the vehicle and a system that could handle the steering, gas pedal, and brakes independently. Another piece of the puzzle was a graphics processing unit that took all of the information that the camera saw and made sense of it. The algorithm Wayve utilized was a model-free deep reinforcement learning algorithm that was used to solve the lane following task. With all of the necessary components on the vehicle, the algorithm worked through three processes: exploration, optimization, and evaluation. 

Wayve Autonomous Car.jpg

How It All Works 

Exploration, as Wayve states, refers to the algorithms ability to "try different actions and record data." Optimization is the process of updating the model using the available data, while evaluation sees "how well the model performs." 

To see if everything worked, the company put the little Twizy on a curvy – not one with too many corners, though – road and let it go. A human driver was present behind the wheel of the vehicle throughout the entire test and took control of the machine when it started to swerve. 

As the company's three-headed approach reveals, the Twizy "experimented," making mistakes and learning about its surroundings through random actions. When the system made a mistake, the human driver was there to intervene, straightening out the vehicle and putting it down the right path before letting go of the steering wheel again. 

Every time the system traveled a long distance, it was rewarded. The human safety driver optimized the policy from within the vehicle, ensuring that the system understood what it did well. 

Wayve's System Impresses In A Packed Field

As you can see from the video below, the system learns quickly. It goes from making numerous mistakes to becoming smoother and smoother, acting more like a regular human driver as it learns. After 20 minutes, which has been condensed in approximately two minutes in the video below, the vehicle learned how to stay on the single-lane road on its own. 

For Wayve, it's all about the quality of the AI systems that are placed in autonomous vehicles instead of the quantity.

"DeepMind have shown us that deep reinforcement learning methods can lead to super-human performance in many games including Go, Chess and computer games, almost always outperforming any rule based system," stated Wayve in a blog post. "We here show that a similar philosophy is also possible in the real world, and in particular, in autonomous vehicles." 

The company believes that the semi-autonomous cars on the road today are "stuck at good but not good enough performance levels." With the British company's algorithm, cars can go from okay to really good over a short period of time. As the company puts it, a driving algorithm that starts with operating at roughly 95 percent of a human could quickly become better. 

How Wayve's System Can Make Autonomous Cars Better

"After a full day of driving and on-line improvement from human-safety driver take over, perhaps the system would improve to 96 percent," states the company in the blog post. "After a week, 98 percent. After a month, 99 percent. After a few months, the system may be super-human, having benefited from the feedback of many different safety drivers." 

The whole goal from the get-go with autonomous vehicles is to make them safer than human drivers, but to operate in a similar fashion. While other automakers are moving towards the goal with numerous sensors, cameras, and computers, Wayve's approach is different. 

"Wayve has a philosophy that to build robotic intelligence we do not need massive models, fancy sensors and endless data," claims the company in the blog post. "What we need is a clever training process that learns rapidly and efficiently, like in our video above. Hand-engineered approaches to the self-driving problem have reached an unsatisfactory glass ceiling in performance. Wayve is attempting to unlock autonomous driving capabilities with smarter machine learning." 

Other companies, like Tesla, use learning elements to make their machines better. Tesla's Autopilot system makes a note in its system every time a driver makes a mistake, i.e. a driver has to take over from the system, and spreads that information to other Tesla's so that it doesn't happen for other drivers. Wayve's idea of letting the vehicle learn on its own, though, is innovative, bold, and fascinating, which is exactly what the industry needs to bring fully autonomous vehicles onto the road. 

Prev                  Next
Writer's other posts
    Related Content