Serkan Arslan, director of automotive for Nvidia, gave the second presentation of the Autonomous Vehicle Software & AI Symposium, demonstrating the sheer complexity of developing and validating the software that powers AVs and highlighting some of the issues Nvidia has seen with the companies it assists in the AV sphere.
“One of the things we’ve learnt in working with lots of startups is that a proof of concept can be done really quickly, but transferring that to the ‘automotive brain’ is probably about 90% more work.”
“The mapping companies have a huge asset of data, but there’s a bottleneck in trying to process it all to output HD maps. The even more challenging part is trying to make sure the maps are up-to-date, sometimes in real time.”
“Sensors are another issue, because perception is based on what the cars are sensing. There are a lot of philosophical discussions about whether we need lidar, radar or cameras. Nvidia’s approach is that you should use all of them, and then optimize the system when it becomes clear you don’t need a certain sensor.”