U.K. self-driving car startup Wayve, a company founded by Amar Shah and Alex Kendall, a pair of artificial intelligence (AI) Ph.Ds from the Cambridge University, believes that it has found a better way than the existing techniques of sophisticated hardware and labor-intensive 3D maps to train autonomous vehicles. On Monday, the company released a video in which a modified Renault Twizy teaches itself by using Wayve’s model-free deep reinforcement learning algorithm to follow a lane from scratch in 15 to 20 minutes using nothing but a computer and a single camera.
The young company, which details its experiment in a paper published on arXiv.org, believes not only in no hand-engineered approaches to the self-driving problem but also in the idea that one efficient way to making self-driving cars capable to drive safely in any environment is that of focusing on the self-learning capability aspect of the car’s software. And that`s why Wayve through its no cloud connectivity or use of pre-loaded maps system is creating the world’s first autonomous cars entirely on deep-reinforcement learning.
“The missing piece of the self-driving puzzle is intelligent algorithms, not more sensors, rules and maps,” says Shah, Wayve co-founder and CEO. “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 approach in mind, Wayve’s team which is made up of experts in the fields of robotics, computer vision and AI from both Cambridge and Oxford universities equipped their Twizy with only one camera. That one camera was hooked it up to Wayve’s four layer convolutional neural network that performs all of its processing on a GPU inside the car. The GPU ran Wayve’s algorithm, which controlled the vehicle’s acceleration, braking, and steering. As you will see from the video posted below, a human driver sat behind the wheel letting the car experiment with the controls.
Every time the car veered off the road, they stopped it and corrected it. The algorithm “penalized” the car’s system for making mistakes, and “rewarded” it based on how far it traveled without human intervention. Within about 20 minutes, which took less than 20 trials, the car figured out how to follow the gently curving road indefinitely.
While a lot of work remains to be done before Wayve’s algorithm can drive safely in any environment and under any circumstances, according to the start-up, these algorithms are only going to get smarter:
“Imagine deploying a fleet of autonomous cars, with a driving algorithm which initially is 95% the quality of a human driver. Such a system would not be wobbly like the randomly initialised model in our demonstration video, but rather would be almost capable of dealing with traffic lights, roundabouts, intersections, etc. After a full day of driving and on-line improvement from human-safety driver take over, perhaps the system would improve to 96%. After a week, 98%. After a month, 99%. After a few months, the system may be super-human, having benefited from the feedback of many different safety drivers.”
Reference: New Atlas