Pittsburgh's Edge Case Research can simulate self-driving car 'nightmare scenario'

A pair of Pittsburgh companies have teamed up to design software they hope will make it safer to test self-driving cars in the city and around the world.

Simulation software developed by Edge Case Research and detailed 3D maps of Pittsburgh gathered by Kaarta technology will allow companies developing self-driving cars to test vehicle software and artificial intelligence in countless scenarios without putting a car on the road.

Mike Wagner, co-founder of Edge Case Research, said the company's Hologram simulation software can tweak lighting and weather conditions or add chaotic traffic or unpredictable pedestrians to mapping data. As self-driving cars roam a city, they collect detailed maps of streets, sidewalks, buildings and intersections. Hologram enhances the complexity of the situations in that mapping data. The software can run a self-driving car through millions of scenarios in a matter of hours.

“It turns it into sort of a nightmare scenario,” Wagner said. “We take each of those miles driven, and we add in everything that a car might see.”

Edge Case announced the release of Hologram on Wednesday. The company has five tests of the software running with companies developing self-driving cars. Wagner would not say which companies are testing Hologram.

Simulation testing is a way for autonomous vehicle companies to build up virtual miles to see how its perception software, artificial intelligence and deep learning algorithms handle situations. It can take months or years for fleets of self-driving cars to build up millions of miles let alone the billions of miles that some experts say autonomous vehicle companies will need to demonstrate their systems are safe. Waymo, Google's self-driving car company, just reached 5 million miles. Uber hit 2 million miles in late 2017. Nvidia announced at the end of March that its new simulation technology can simulate about 1.5 million miles in a day.

Hologram makes those millions of miles extremely tough miles for the car's systems, Wagner said.

“Machine learning works by example, so the more examples you give it, the more robust it should be,” Wagner said.

Calls for more simulation testing of autonomous vehicle software, artificial intelligence and deep learn algorithms have grown after a self-driving Uber last month hit and killed a woman walking her bike across a street in Tempe, Ariz. Uber has suspended its testing of self-driving cars in Tempe, Pittsburgh, Toronto and San Francisco as the company, Tempe police and federal authorities investigate. It remains unclear why sensors and cameras did not notice the victim, Elaine Herzberg, in time for the Uber to avoid hitting her.

Wagner said Hologram was in development long before the Uber crash and that the self-driving car companies he knows and works with do extensive simulation testing. Wagner was at a conference working with others in autonomous vehicles on safety standards when news of the crash broke.

“When there is a large fleet of these things deployed, there are going to be accidents,” Wagner said about the testing of self-driving cars. “But the weight of having that go from an abstract idea to the real thing was hard.”

Wagner said the crash underscored the need for in-depth testing of a vehicle's software. A bug anywhere in the hundreds of millions lines of code could cause a catastrophic failure and lead to a tragic crash.

“That's why people take this stuff seriously and dedicate their lives to it,” Wagner said.

Edge Case worked with Kaarta to create maps of Pittsburgh's Downtown. Kaarta developed Stencil, which combines a LiDAR unit and camera similar to what self-driving car companies use and imaging software to create detailed maps. Stencil can create 3D models in real time and can be used immediately in testing, said Kevin Dowling, CEO of Kaarta.

The mapping data will be available to self-driving car companies looking to simulate tests on the city's streets and to city planners and others who need detailed maps.

Aaron Aupperlee is a Tribune-Review staff writer. Reach him at aaupperlee@tribweb.com, 412-336-8448 or via Twitter @tinynotebook.