Virtual programs help 'teach' autonomous cars how to drive

Moving down a city street, the car brakes suddenly as a police cruiser bursts across the road.

Adapting quickly, the car adjusts and continues to trundle along within the speed limit, diligently observing traffic laws and reacting to surrounding vehicles.

The car is not really there and the environment – the road, the pavement, other cars and pedestrians – are all virtual. The video game-style program is a “driving school” for a new autonomous vehicle, one of the most innovative topics this morning at the Fisita Plus conference in central London.

Simulated systems will partially “train” commercially available self-driving cars in the near future, said computer vision professor Dr Luc van Gool. The main thing enabling self-driving cars is data, said the professor – “masses and masses” of it. Camera and control systems in autonomous vehicles must be extensively tested in a huge variety of situations to ensure they are safe and efficient, and simulations provide a completely customisable and adaptable source of training scenarios.

“You can take the exact same scenario and scene with certain levels of noise and rain for example, and compare,” said Dr van Gool to Professional Engineering. “You can’t do that in real circumstances… Performance goes up under the influence of this additional data.”

However, he stressed virtual training can only supplement real-world scenarios. “We need to stress that before we have headlines about cars being ‘Trained by Pacman’,” he said. “It all depends how the cars do on the roads.” He refused to estimate how soon a partially simulation-trained car could be commercially available.  

Some companies are already using simulation to quicken the autonomous revolution. London-based Immense Simulations exhibited at Fisita Plus, showing how they create city-sized simulations to predict and test demand for transport companies.

The start-up’s simulations are also used to virtually test vehicle sensors or examine how driverless cars might interact with traditional vehicles. Simulations provide a safe and easy way of testing new models before they are introduced to real-world scenarios, said Immense Simulations expert Vittoria Parisi.

“Simulation allows you to test different strategies and understand which one is the best one,” she said to PE. “It is a preliminary step to understanding what the future will look like.”