Chinese and American researchers have developed a data-driven simulator that could further improve the safety of autonomous cars, as self-driving vehicles will have to pass tests, just like humans, before they can hit the road.

The study published earlier this week in Science Robotics, a scientific journal, showed that the new photo-realistic system could provide a more vivid and authentic simulation than current systems that mainly rely on gaming engines and high-fidelity computer graphics.

Scientists from China’s search engine Baidu, the University of Hong Kong and the University of Maryland developed the novel system called Augmented Autonomous Driving Simulation that could make self-driving technology easier to evaluate so as to iron out any deficiencies inside a lab, before an actual road test.

The current simulator technology widely used by researchers and automakers uses a perception module, in which a self-driving car receives input from computer-generated imagery with mathematically modeled movement patterns to mimic pedestrians, bicycles and other road users.

The major drawback of the current system is that it is a relatively simplified representation of real scenarios on the road, and cars and self-driving software that pass the “mock” test may not be necessarily safe to be used in the real world.

A Baidu autonomous navigation system on display. Photo: Handout

The new system combines photos, videos and lidar point clouds with real-world trajectory data to simulate pedestrians, bicycles and other cars. Those trajectories can be used to predict the driving behavior and next positions of other vehicles and pedestrians for safer navigation.

With the new system, vehicles and pedestrians can also be “lifted” from one scenario and “placed” into another with adjustable lighting and movement patterns to simulate real-world road conditions.

“The way humans drive is not easy to capture by mathematical models and simple, theoretical laws of physics. So we extracted data about real trajectories from all the videos we had, and we modeled driving behaviors using social science methodologies and a new set of algorithms,” said the paper’s co-author Dinesh Manocha, a computer scientist with the University of Maryland.

“With a simple scan, the new system can automatically produce realistic car flows and related data at a much lower cost,” the paper’s co-author Li Wei with Baidu Research told Xinhua.