A new U.K. self-driving car startup founded by Amar Shah and Alex Kendall, two machine learning PhDs from University of Cambridge, is de-cloaking today. Wayve — backed by New York-based Compound, Europe’s Fly Ventures, and Brent Hoberman’s Firstminute Capital — is building what it describes as “end-to-end machine learning algorithms” to make autonomous vehicles a reality, an approach it claims is different to much of the conventional thinking on self-driving cars.
Specifically, as Wayve CEO Shah explained in a call last week, the young company believes that the key to making an autonomous vehicle that is truly just that (i.e. able to drive safely in any environment it is asked to), is a much greater emphasis on the self-learning capability of its software. In other words, self-driving cars is an AI problem first and foremost, and one that he and co-founder Kendall argue requires a very specific machine-learning development skill set.
“Wayve is building intelligent software to decide how to control a vehicle on all public roads,” he tells me. “Rather than hand-engineering our solution with heavily rule-based systems, we aim to build data-driven machine learning at every layer of our system, which would learn from experience and not simply be given if-else statements. Our learning-based system will be safer in unfamiliar situations than a rule-based system which would behave unpredictably in a situation it has not seen before”.
To explain his thinking in laymen’s terms, Shah points to the way a human who is relatively proficient in driving in one city can quickly adapt to the differences in a completely new city, without having to be given extra training or instruction beforehand. It may take around 30 minutes or even a few hours to become fully climatized to new driving conditions or environment, but humans don’t need very much new data to do so.
“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,” he says. “We want to give our vehicles better brains, not more hardware”.
The problem, thus far, the pair argue, is that companies like Google and Uber are throwing an engineering mindset at making vehicles autonomous, in the sense of designing rule-based systems that try to pre-empt and deal with every edge case, whilst in tandem adding more sensors and capturing more data. This might produce encouraging results in the specific, narrow setting it has been engineered for, but won’t have maximum payoff longer term.
“Right now, big tech companies have cars with many different sensors of a handful of different types. Their attitude is to have more and more sensors to do more and more difficult driving tasks,” says Shah. “If I ask you to do a difficult athletic obstacle course, something like Ninja warrior, having more eyes isn’t really going to help you much. What you need is better coordination – it’s the mind-muscle connection that’s the limiting factor. In driving, it’s really the way you use your sensory information that’s key (the AI-wheel connection in the car), not the number of cameras and radars and LIDARs”.
But if a more sophisticated machine-learning approach is the correct one, surely Google (which has several AI efforts under its parent company, including being the owner of DeepMind), would already be going down that avenue, too?
“The big teams are distracted by getting something working because they have stakeholders who have been investing for a decade into autonomous driving. They are getting impatient,” the Wayve co-founder pushes back. “How will Alphabet tell their shareholders ‘we’ve invested X billion USD into Waymo and its predecessor with a team of 1,000s, but we are now throwing that approach all down the drain and hiring more AI people to solve driving’. It’s a hard sell having spent billions and when they are close to a simple product. Same reason politicians make bad long-term decisions… their output is only short-term”.
“Wayve has a very differentiated technical approach versus most other autonomous vehicle startups,” echoes Fly Ventures’ Gabriel Matuschka. “It’s a 10x improvement over the rules-based approach taken from legacy robotics to hard-code the driving actions that the vehicle takes once it understands what it sees. Wayve uses end-to-end machine learning to drive cars autonomously, with little data, in novel environments. This means that their software enables a car to drive itself using only understanding of what it can see, just like humans do”.
To that end, the ten-person Wayve is said to be made up of experts in robotics, computer vision and artificial intelligence from both Cambridge and Oxford universities, who have previously worked at the likes of NASA, Google, Facebook, Skydio and Microsoft. Their work ranges from using deep learning for visual scene understanding to autonomous decision-making in uncertain environments. Noteworthy also is that Professor Zoubin Ghahramani, Chief Scientist of Uber, is an investor in Wayve.
“There are very few teams out there with the academic background and technical capabilities to at all have a credible shot at this. Wayve is one of them,” adds Matuschka. “Some people in the industry question if Wayve’s novel approach will work. You only stand a chance to compete against Google, Uber, et al. if you try, and are able to do something that the large players haven’t done so far or don’t believe in yet. Then you can have a head start”.