The car learns to drive like we do — through trial and error.
TRIAL AND ERROR. Artificial intelligence startup Wayve believes it's found a better way to train autonomous vehicles (AVs) than the existing methods of sophisticated hardware and detailed 3D maps. On Monday, the company released a video in which a modified Renault Twizy (a two-seated electric vehicle) learns to autonomously navigate a road. It does so through reinforcement learning, a type of machine learning in which a system earns a "reward" for desirable behavior and a "penalty" for undesirable behavior. They detail their experiment in a paper published on arXiv.
THE PROBLEM WITH MAPS. Most in-development AV systems rely on detailed 3D maps to navigate. Companies across the globe are currently racing to create these maps, using sophisticated sensors and cameras to detail city streets and highways alike. AVs, in turn, require their own complex systems of cameras and sensors to then navigate these maps.
Unfortunately, 3D maps are labor-intensive and require frequent updates to account for things like construction. Companies that create 3D maps also tend to focus on highly trafficked roads first, leaving rural areas behind.
TEACHING A TWIZY. Wayve's technique could render those 3D maps obsolete. The Twizy in the video had only one camera mounted at its front (most other AVs have many more — Tesla's autopilot, for example, uses eight). That one camera fed information in real-time to a graphics processing unit (GPU) onboard the car. The GPU ran Wayve's reinforcement learning algorithm, which controlled the vehicle's acceleration, braking, and steering. A human driver sat behind the wheel during the learning process, stopping the car every time it veered off the road, which "penalized" the system. The longer the system drove without this human intervention, the greater the "reward" it received.
Within about 20 minutes, the car figured out how to follow the gently curving road indefinitely.
If Wayve's approach to autonomous driving catches on, AVs would no longer need to rely on 3D maps. And that could mean that anyone, anywhere could enjoy the benefits of autonomous vehicles.
READ MORE: AI Algorithm Teaches a Car to Drive From Scratch in 20 Minutes [New Atlas]