A Self-Driving Driving License

Safety experts at TÜV SÜD and AVL are working with NVIDIA to use DRIVE Constellation for the validation of future autonomous vehicle standards, says Zvi Greenstein.

TÜV SÜD, a safety agency that works with governing bodies to draft vehicle regulations, in collaboration with AVL List GmbH, has created an initial set of proposed testing measures for manufacturers’ automated driving systems, running them on the NVIDIA DRIVE Constellation simulation platform.

AVL combines the real and virtual worlds in various stages of automotive development. In addition to purely virtual and hardware-in-the-loop testing, AVL builds the bridge to testing real vehicles by running the physical vehicle on a test bench for autonomous driving called AVL DRIVINGCUBE™.

These simulation metrics developed by TÜV SÜD evaluate how automated driving systems will behave in regular traffic scenarios before they go on to public road testing.

While traffic accidents happen everyday, it’s rare for an individual driver to see one — in the U.S., one accident occurs per 492,000 miles driven, according to federal data. However, autonomous vehicles must be prepared to avoid accidents before they drive in traffic.

To do so, manufacturers and regulators are working on ways to validate self-driving technology, ensuring it will behave as intended while operating. With highly accurate simulation, governments can begin to put regulations to place and test these standards in the virtual world — enabling them to cover these relatively rare, yet significant traffic hazards, establishing comprehensive safety practices in deploying autonomous vehicles.

A No-Stress Driver’s Test
DRIVE Constellation is a hardware-in-the-loop simulation solution that allows manufacturers to test autonomous driving technology with unprecedented accuracy. The platform uses two servers. One runs DRIVE Sim software, which generates a photorealistic driving environment from real-world sensor data. The other contains the DRIVE AGX Pegasus compute platform, the exact hardware and software configuration that would run in the car.

The bit-level accuracy enabled by using both the identical software and hardware configurations in simulation allows regulators to closely mimic the vehicle dynamics and traffic scenarios of the real world before the rubber meets the road. The way a car reacts to a sudden stop, low visibility in inclement weather conditions and unexpected road hazards can all be reproduced in DRIVE Constellation with the highest levels of fidelity.

TÜV SÜD’s initial testing scenarios cover how an automated driving system handles common highway traffic situations using a set of defined performance metrics. For example, the simulation would test whether a vehicle sufficiently slows when another car cuts in front of it, then speeds back up once traffic moves forward again.

Like a driver’s license test, these cases don’t just evaluate whether a car can drive without crashing, but also how safely and comfortably it can handle everyday situations. To do so, TÜV SÜD uses performance indicators such as time to collision — the number of seconds it would take for a test vehicle to collide with the car in front of it, if brakes weren’t applied — and how well the car maintains a safe following distance. These metrics are tracked at every point throughout the test cases.

Testing in the Future
By pairing simulation with on-road testing, regulators can develop a robust testing process. Building off of this first set of test scenarios on DRIVE Sim, TÜV SÜD plans to expand evaluations in the virtual world to city driving — like making a left turn at a crowded four-way intersection — and testing the car’s lateral dynamics, such as how well it handles tight curves, lane changes and other side-to-side movements.

As with all of the safety agency’s tests, these proposed cases were developed for widespread adoption across manufacturers and global regulatory bodies. And, by working with NVIDIA DRIVE Constellation, these proposed standards can continue to evolve to cover even more of the rare and difficult scenarios AI drivers must be able to handle for truly safe deployment.

This collaboration is just the beginning of joint efforts between NVIDIA, TÜV SÜD and AVL to develop virtual testing standards as part of the autonomous vehicle validation process worldwide. As the industry advances toward a self-driving future, this type of cooperation will ensure the safety benefits of autonomous vehicles will be fully realized.

Author