Adopting an empirical approach and shunning methods based on hardware and detailed calculations, two British engineers have created a machine learning-based system that can teach a car to navigate a road in only 20 minutes. The two have set up startup called Wayve and are looking to improve the algorithm’s performance by develop its learning capabilities rather than fit equipment into cars.
Wayve, a startup founded by two Cambridge University students, has developed an AI-based system capable of teaching an autonomous car to drive in just 20 minutes. The algorithm has developed the ability to control the route taken by a car without it leaving the road, as reported by Futura Sciences. The short learning time is all down to the innovative approach taken by the engineers.
A speedy learning process
Most existing techniques for training autonomous vehicles involve a significant number of calculations and the use of data transmitted by various devices. The cameras, sensors and LiDAR technology fitted in connected cars allow them to map and follow a route and avoid obstacles. The process is long, however. In contrast, Wayve’s AI teaches the car by following an empirical experimentation process that mirrors the way in which humans learn to drive.
The key to this latest development is reinforcement learning, the same method used by DeepMind in developing AI capable of playing the Quake II Arena video game in a team. Wayve’s engineers have harnessed it to activate a multi-layer neural network and thereby create a much faster system, one in which only 20 attempts were needed before the car remembers how to follow the road.
The test phase involved a driver sitting at the wheel of a Renault Twizy and equipped with a camera. Whenever the programme caused the car to leave its desired path, the driver stopped the car. Thanks to the corrections made and the car’s growing ability to drive further and further without the need for human intervention, it quickly learned from the mistakes it made.
“We want to give our cars better brains, not more equipment,” said the project leaders. Now that their algorithm is capable of navigating a route, Wayve’s engineers are now looking to go even further. The next phase involves teaching the car how to negotiate junctions, roundabouts and traffic lights in complete safety.
Contact Allianz Partners
Jul 11, 2018
The diversity of technologies developed and variety of terminologies used by automotive manufacturers to advertise their smart cars can lead to some degree of confusion. To give you som [...]
Jan 16, 2018
Nissan has announced that its driverless taxis are to undergo tests in actual road conditions in Yokohama, Japan next March over a two-week period.