UST 031

EasyMile EZ10 autonomous shuttle | Dossier developed by a separate team runs on another. This is a method drawn from the aviation industry, in which redundant flight control software is written by different teams to minimise the risk of both systems having the same faults. Development challenges Despite the availability of GPS and Lidars, localisation in open areas proved a major challenge, Pairot says. The effectiveness of Lidars is limited in flat areas, he adds, because their signals are not reflected back to the transceiver from every element of the environment that they need to be, in order to build the 3D picture the system uses to understand and navigate the environment. “We needed to create software intelligent enough to mix all this data together and make it work in any situation,” he says. From a safety point of view, the overarching challenge has been to ensure that the EZDrive software can understand what is happening around the vehicle and to adapt its behaviour accordingly. Early in the development programme, leaves falling in front of the vehicle would be detected by the Lidars and would trigger an emergency stop (‘e-stop’) – in one incident, in Australia, a particularly large butterfly did the same. Now the system is much more refined, and most e-stops are for genuine emergencies, but work continues to find the ‘corner cases’ – extreme conditions that combine multiple ‘what-ifs’ to find out what the vehicle will do in unlikely but possible circumstances that might reveal bugs or risky behaviours – and train the algorithms to respond appropriately. Pairot emphasises that false positives which trigger a stop decision for no good reason are greatly preferable to false negatives that might lead to a decision not to stop the vehicle Some authorities asked us to remove the bidirectional features because they don’t allow beams from the rear of a vehicle