Aurora, the autonomous vehicle startup scheduled to make its debut on Nasdaq next week, is turning to a team of Pixar veterans to help make the computer simulation tool used to test and train its self-driving system more like the real world.
The three-person computer graphic imagery startup Colrspace, which has been operating in stealth, is joining Aurora's perception team. Aurora will also own Colrspace's IP, specifically technology that combines CGI and machine learning. The trio, which includes Michael Fu, Allen Hemberger and Alex Harvill, has developed a technique that can reconstruct 3D objects and materials from a photo or image. Essentially, it makes the simulation more "realistic," which Aurora and other autonomous vehicle developers contend helps make testing more effective. (Below is a sample of Colrspace's work; TechCrunch turned the mp4 into a gif)
Image Credits: Aurora
Fu, Hemberger and Harvill will join Aurora's perception team, which already includes people from 7D, a simulation startup founded by former Pixar software engineer Magnus Wrenninge. TechCrunch learned in 2019 that it had acquired 7D.
While Aurora and competitors like Argo AI, Cruise and Waymo regularly conduct real-world tests on closed courses and public roads, computer simulations are seen as an essential tool to test, train and validate their autonomous vehicle technology. Simulation can be used to help the self-driving system test various scenarios or play back what happened in the real world. The end game is to train and evaluate the software stack so it's safe in the real world.
It's not uncommon for the biggest AV companies to run through thousands, even millions, of simulations daily. Aurora, for instance, estimates its simulator is the equivalent of running more than 50,000 trucks driving continuously. Aurora's computer simulator, called the Virtual Testing Suite, runs through a diverse range of driving conditions as well as common and unusual (edge case) scenarios to catch errors early and before it is deployed in vehicles operating on public roads. Data captured while driving on public roads is also fed back into the simulator.
This year, Aurora has put more effort and resources to scale up its simulation program. Aurora said this month that it expects to have driven the equivalent of over 9 billion simulated miles by the end of the year, with 6 billion of these miles clocked so far in 2021.