Driverless cars are already cruising around San Francisco. Soon they will be part of a public ride-hailing service, thanks to Waymo. The company is a spinoff of Google’s self-driving car project, which has been testing cars in the city for over a decade. Waymo vehicles are also taking millions of photos. From these, the company has rendered an entire neighborhood in 3D. This simulation can help train the next generation of self-driving cars.
Waymo calls the technology Block-NeRF. It is an extension of the NeRF (neural radiance fields) technology that uses artificial intelligence to make 2D photos 3D. The neighborhood recreation using photos taken by Waymo’s self-driving cars is by far the largest scene rendered to date using this technology. Waymo already uses a virtual environment for training its self-driving technology. A fleet of virtual cars has driven more than 20 billion miles on top of the more than 20 million miles driven by its cars in the real world.
Waymo posted a research paper and videos of the recreations on their website and the academic site arXiv. The dataset includes photos from various times of day and weather conditions. Stationary objects are well rendered. But moving objects like foliage ends up blurred due to motion and changing seasons.
Cars, bicycles, pedestrians, and other potential traffic hazards are not yet part of the recreation. The next step for this virtual world is to speed up the rendering process, ideally making it real-time. The goal is also to sharpen distant objects.
Waymo announced that one of its fleet of Jaguar SUVs drove a company engineer to work without a driver. It will continue to be tested with employees and then opened to the public. Waymo’s driverless ride-shares have operated in Phoenix, Arizona, since October 2020.
While it will be disconcerting to see a driverless car, it’s likely much safer than the alternative. Distracted drivers are much more likely to cause accidents than artificial intelligence.