Algorithm created by British start up teaches cars to drive themselves
Endeavours to create autonomous cars are nothing new. Google’s self-driving car project began in 2009 and, by 2012, had racked up over 300,000 miles of self-driving accompanied by test drivers. Since then, the focus has been on full autonomy, with numerous tests proving successful. But now, a British start-up has found a way to teach cars to drive themselves using AI.
Called Wayve, the start-up claims to have found a better way than existing methods to train autonomous vehicles. The current preferred method uses sophisticated hardware and detailed 3D maps which, while efficient, are labour-intensive and require frequent updates to account for changes in road systems.
A machine learning solution
Wayve has taken a different approach. Last month, it released a video showing a modified Renault Twizy learning to navigate a road autonomously. The machine learning solution used to achieve this is called ‘reinforcement learning’ and delivers a ‘reward’ to the system for desirable behaviour.
The approach does away with much of the time-consuming and complicated technology, enabling a car to self-drive with only an understanding of what it can ‘see’ – like humans. The company claims that vast amounts of new data and dense mapping technology will not be necessary with its solution.
The Twizy test
In the video, the Twizy uses a single front camera to feed information to a graphics processing unit (GPU), which runs Wayve’s reinforcement learning algorithm. This algorithm operates the Twizy’s acceleration, braking and steering and is accompanied by a human driver, who corrects it when it goes off course. These corrections ‘penalise’ the system, causing it to learn how to follow the route; the longer it drove without human intervention, the greater the ‘reward’ delivered. In total, it took around 20 minutes for the car to understand how to follow the curving country lane.
Taking the technology further
Wayve has already cut down the amount of technology needed for the autonomous driving experience. The Twizy’s one camera makes other AVs seem over-complicated, with models such as Tesla’s Autopilot relying on no fewer than eight cameras. There’s much more to be done, though.
Wayve is now working on scaling the reinforcement learning technology to accomplish more complex driving tasks. It’s using the idea that human brains can learn in minutes what some algorithms take millions of attempts to perfect, and seeks to provide a solution that solves tasks quickly and effectively.
The innovative start-up, which was established in January 2018, is undoubtedly eligible for R&D tax credit for software companies. In a paper detailing the experiment, the founders invited other businesses to follow their example and explore the possibilities of reinforcement learning in connection with autonomous driving.
Organisations looking to take up the challenge would also qualify for UK R&D tax credit; those who are already investing in software development for self-driving cars or reinforcement learning techniques for other purposes can use our R&D tax calculator to find out how much remuneration they could claim for their work. Contact us today to find out more.
