Tesla, Uber, Google’s Waymo are just some of the big names that experiment with self-driving vehicles. Apple is another company that joins them in such endeavor after it received a permit on Friday from the California DMV that allows them to test autonomous vehicles on the state’s roads.
Apple doesn’t want to talk about future projects and products, so we don’t know much about it. According to Apple’s spokesperson, the company started focusing on “machine learning and autonomous systems.”
It was two years ago when we heard rumors about a tech giant being interested in autonomous vehicles. They were restrained from commenting at first, but the signs that indicated that developers were working on self-driving software were everywhere. The new patent that Cupertino-based company filed is called “Collision Avoidance of Arbitrary Polygonal Obstacles” and it is about the techniques that “use simple geometry to identify which edges of which obstacles an agent is most likely to collide.” The goal for a vehicle is to see what objects it could possibly crash into and avoid the hit by calculating the “avoidance force.”
What is key for accelerating testing of self-driving cars is a simulation, and Waymo’s self-driving software has reached more than a billion miles on simulated roads. There is a possibility that Apple also spent hours testing the new software in a simulated environment.
Its first and only research paper so far is about a general adversarial network called SimGAN and its ability to generate image data identical to the real test data in order to spawn more info which will be later on used to further train its models.
Considering that SimGAN produces volumes of fake, realistic labeled data which can confuse the systems and make it think it’s real data, a major problem can be avoided. The issue is that ongoing systems need huge swathes of data which need to be labeled carefully. It is a lucrative business, and many startups are stepping up to provide training data as a service.
It is possible that Apple is going to use SimGAN to enhance its autonomous car efforts, and this has been indicated in the report conclusion. “In [the] future, we intend to explore modeling the noise distribution to generate more than one refined image for each synthetic image, and investigate refining videos rather than single images,” the paper says.
