Toyota Invests in CARLA Open Source AV Simulator Project

Toyota’s autonomous-vehicle division in Silicon Valley has invested $100,000 in CARLA, an open source simulation project, to take advantage of the collaborative development potential of an open-source community.

The investment by Toyota Research Institute (TRI), announced on June 21, is a drop in the bucket compared with the parent company’s research and development budget, which topped $9 billion last year. In fact, it wouldn’t even buy CARLA one limited-edition Lexus LC Inspiration Series sports car (starting price $108,180).

Toyota is investing in AV development from several angles. Among other things, it recently built an exclusive closed-course test site in Michigan and invested $2.8 billion, along with partners, in a new software company in Tokyo. But open source can provide one thing that big bucks won’t necessarily buy, which is data and expertise from a diversity of sources.

TRI notes that the CARLA (Car Learning to Act) will help it to exchange code, information and data with industry and academic partners. The project, administered by the Computer Vision Center (CVC) in Barcelona, counts Intel as its founding sponsor. CARLA includes open source code and protocols as well as digital assets, like virtual buildings, vehicles, traffic lights, trees and pedestrians, that users can integrate into simulated environments. One recent addition is a simulated Tesla Model 3.

It also includes tools for simulating different weather and lighting conditions, plus virtual sensors such as cameras and GPS devices that can be applied to virtual vehicles.

The code and resources are available free of charge on GitHub, making them easily accessible to developers around the world. CARLA is focused on urban driving, which presents some of the most complex scenarios for AVs to master. It can be used for prototyping, training and validating driverless car platforms, CVC says.

Developers can use a standard development model based on perception, planning and control, a learn-by-imitation model, or learning by reinforcement, as demonstrated in a video that shows a virtual AV breaking all the rules until it crashes.

Simulators like CARLA are increasingly important in the AV industry, especially since the fatal crash in March involving an Uber test vehicle in Arizona. TRI halted its own tests on public roads after that accident, as did Uber.

Testing in simulation allows virtual vehicles to log billions of miles of driving that just wouldn’t be possible on real roads with human drivers. It also allows AV systems to go through particular scenarios repeatedly and under different conditions, including simulated weather and lighting. Graphics computing and AI vendor Nvidia has invested heavily in simulation platforms for AVs, which can be integrated with its computing engines for driverless cars.

One of the best ways the AV industry could use open source is to gather data about real-world accidents, University of Notre Dame autonomous systems expert Timothy Carone told The Connected Car.

“All these different kinds of accidents that occur in the US could be put into a library … and they’re realistic. They’re accidents that really happened,” Carone said. Sources might include police departments, insurance companies, regulatory agencies and even auto manufacturers, he said.

That’s smarter than a company developing its own simulator and trying to postulate all the possible scenarios that AVs might face, he said: The real world knows better.

— Stephen Lawson is a freelance writer based in San Francisco. Follow him on Twitter @sdlawsonmedia.