To be safe in the real world, AVs must spend time in a virtual one

To ensure the safety of autonomous vehicles, companies have been testing fleets in San Francisco, Austin, Miami and elsewhere — gathering data and exposing their technology to everyday experiences. However, road tests are a cumbersome form of validation — the Rand Corporation estimates it would take hundreds of millions to hundreds of billions of miles (nearly a century of driving) to prove an AV drives safely.

The big picture: Not all experience needs to come from road tests. Simulation platforms enable the artificial intelligence brain powering an autonomous vehicle to run in a photorealistic world that mimics real-life traffic, exposing its deep-learning algorithms to scenarios and conditions as many times as necessary for the system to handle them perfectly.

Manufacturers can also test hardware using a process called hardware-in-the-loop, in which one server simulates the driving environment while another contains the computer that will eventually run in the car.

• The output of the first server is the input to the second, running the same software as it would if it were driving on the road.
• The commands to control the vehicle are then sent back to the simulator 30 times every second, completing the "loop" and allowing engineers to test the software in the same hardware environment as it would operate in the vehicle.

Human-in-the-loop testing adds another layer.

• Engineers can take over other vehicles in the virtual environment — like cars in a video game — driving them around the AV and testing how it reacts to sudden cut-offs or tailgaters.
• When combined with public road testing, this type of simulation creates a robust validation process in a fraction of the time it would take to drive an equivalent distance in the real world.

Yes, but: Though virtual worlds like those seen in video games and movies can closely mimic the look and feel of reality, their scripted scenarios lack the same fidelity when it comes to the real world’s diversity and unpredictability.

• Drawing datasets from simulation partners, regulatory agencies and other sources onto a single platform is crucial for comprehensive testing. With an open platform, third-party developers can provide countless cityscapes, vehicle dynamics models and traffic scenarios.
• The same situation can be run over and over again, with a nearly infinite number of changes each time — from time of day, to weather, to an animal crossing the road. And with the addition of human-in-the-loop testing, traffic scenarios can be improvised on the spot.

The bottom line: While scripting the infinite number of potential traffic situations may be impossible, enabling diversity and spontaneity in simulation is a vital way to test a car’s reaction to unforeseen scenarios, validating AV technology without sacrificing safety.

Danny Shapiro is senior director of automotive at NVIDIA.