Unmanned Systems Technology 020 | Alpha 800 I Additive Manufacturing focus I USVs insight I Pegasus GE70 I GuardBot I AUVSI Xponential 2018 show report I Solar Power focus I CUAV Expo Europe 2018 show report

20 In conversation | Chris Hoyle create appropriate models of the world as those sensors would see it, to ensure that the virtual system as a whole reflects real-world interactions between the vehicle and its environment. Monster radar That means scanning the world with the same kinds of sensors. That is feasible with Lidar and video cameras, but Hoyle points out that there is a major challenge with radar modelling of the real world. The problem is the amount of energy involved. “With Lidar scanning you can do an absolutely outstanding job and remain eye-safe, whereas we have done some ‘back of the envelope’ calculations and we think the radar scanner you would need to use would fry you,” he says. “We are talking about measuring the dielectric properties of every object along the side of the road. We do that very well for light, and for Lidar. It is something we aspire to do for radar, but at the moment we don’t think it is technically feasible – you would have to evacuate London and then go in and scan it!” While nothing he has tackled has proved technically easier than expected, he stresses that some of the things he has taken for granted have been perceived as incredibly valuable by new customers from the autonomous vehicle industry, particularly on the AI side. “Being able to physically model a vehicle accurately has been a dream come true for some of those people, yet it was something that we absolutely took for granted as a stable and mature technology,” Hoyle says. The single biggest improvement in simulation technology over the past 10 years, he says, is in the computing power available on graphics cards. Thanks to what he calls an arms race between AMD and Nvidia, their capability has effectively doubled every two years in terms of the achievable resolution, the complexity of materials and subtlety of lighting effects. Customers from motorsport and automotive manufacturing, on the other hand, can generally supply accurate mathematical models of their vehicles for rFpro to wrap the virtual world around. An accurate dynamic model of a vehicle in a high-fidelity virtual world is proving to be a powerful tool with which to teach an AI to drive. “It is critically important that they don’t just create bland, safe AI; they’ve got to create AI that is going to be acceptable to consumers,” Hoyle says. “An out-of-the- box AI has no empathy for its passengers. “We have seen some really good examples of that where the first time an AI company has ridden as a passenger in their vehicle – in a simulator – they have realised that while it might be a very mathematically efficient way to drive, it is jolly unpleasant and makes you sick. As humans, without even thinking, we drive in a way that is pleasant to experience.” That also includes subtle communication with the passengers to ease any anxiety. “It is not enough for your AI to know it is on top of a developing situation; it’s got to give cues to the passengers to let them know that it knows,” he says. The next big hurdle, he says, is to prove that autonomous vehicles can operate safely in the real world, using the virtual parts of it they have created. Hoyle expects customers and test facilities to accumulate databases containing hundreds of thousands or even millions of test manoeuvres, each of which may last only 10 or 15 s. “Those tests will be like regression tests for software, continuously assuring you that your current autonomous model passes them,” he says. In conclusion, he says the testing must prove statistically that the rate at which new failure modes are identified makes the vehicle safer either than all human drivers or at least some carefully measured subset of them. June/July 2018 | Unmanned Systems Technology A computer programmer and car enthusiast from an early age, Chris Hoyle attended the Cass Business School at City University from 1986 to ’89, earning a degree in banking before setting up a series of software companies serving the financial sector. From August 1990 to May 1997 he was technical director of Alternative Business Solutions, developing software to connect mainframe computers to new user interfaces. Between November 1997 and December 1999, he was managing director of two other business software companies, Sigma Computing Associates and SCA Advanced Solutions. He sold his business interests in 2000 to take a sabbatical aboard a yacht, before returning to the UK to start another software r&d company, Docum3nt, at which he was a director from December 2004 to October 2007. In 2006 he gave away an inverse kinematic solver he had created as a hobby project to simulate a Caterham Seven sportscar, which attracted the attention of the motorsport industry and led to the formation of rFpro, where he has served as technical director since November 2008. Chris Hoyle