Autonomous driving requires cars to learn autonomously
SHIFT Mobility recognises the obstacles to driverless cars
Since its 2018 debut at IFA Berlin (then called SHIFT Automotive) SHIFT Mobility continues to demonstrate, explore, explain and debate the future of automotive and mobility technologies. Now that electric vehicles (EVs) are an accepted and essential presence on our roads, the next junction for automotive and tech is autonomous vehicles: when will cars no longer require human drivers? This topic dominated the two day convention.
AI is active in most modern cars and there are several examples of autonomous (driverless) vehicles already operating, for example robo-taxis in designated zones in China and US. Speaking at the conference on Friday, Manuel Yoon, VP of strategy at Israeli software company Autobrains asserted: “The automotive industry has invested $200bn into autonomous vehicle technology. But even Tesla – the most advanced – are not quite there yet”. Yoon suggested that the traditional approach to building the AI and dataset required for safe autonomous driving, where the car “learns” the parameters of the road, is not sufficient. “Manually-labelled data – supervised learning, humans assigning attributes frame-by-frame to captured filmed data – is time consuming, expensive and prone to errors,” he said. “But if you build software that allows the AI to mimic the way humans learn, it removes the need for the human component.” Autobrains has a system: self-learning AI, applying signatures to obstacles/objects on the road, and creating perception fields – imitating how the human brain works. “Driving is intuitive. We can mimic this by creating ‘force fields’ around objects, assigning them likely behaviours based on previous learning,” Yoon added.
The need to adopt unsupervised learning was a common theme. Toby Wessels, chief development officer of Silicon Valley start-up helm.ai, said “Historically, using the traditional method of supervised learning has been a challenge. It can cost a company $5bn and they try to start at L4 [fully autonomous] which is like turning up at the Olympics having never trained for the sport. To achieve autonomy you need three things – the tech, the data and money. Some big companies are not willing to pay – it’s too complex and expensive. We work with OEMs: we license the product to them, with arbitrary input data to teach a system using existing video, and inform AI’s neural nets. At this rate, we should have autonomous driving vehicles by mid-decade, maybe sooner for inner-city driving and fixed repeatable routes – like warehouse to supermarket.”
Greater autonomy leads to greater sustainability
On the topic dominating IFA 2022 – sustainability – the advantages of autonomous driving are manifold. Tobias Wessels, chief development officer, helm.ai, told the SHIFT Mobility conference on Friday: “With no human in the car, there is no need for cooling or heating systems, so the vehicle weight is reduced and less energy consumed. Driving at slower speeds, there is less wind resistance. Often one vehicle can operate for 24 hours, eliminating the need for two vehicles sharing the burden if driven manually. Just these factors can represent a 40% energy saving even over an EV,” Wessels said.
Photo: Self-learning AI-eye-view of the road using the helm.ai technique, courtesy of helm.ai’s self-learning autonomous AI