The first completely autonomous cars will arrive on the market within a few years. Until then, they will have to perfect their driving… among other things in a virtual world, developed by a Quebec company.
Posted at 10:00 a.m.
The LeddarTech company, established in Quebec, has been developing software for several years that allows autonomous vehicles to interpret the environment around them. To do this, their algorithms must gather and merge all the images collected by the cameras and sensors that line the autonomous vehicles. They must also be able to recognize objects important to driving, such as other cars, signs and pedestrians.
However, to develop and perfect these algorithms, which are obviously artificial intelligence, LeddarTech needs to train them with huge databases, which are extremely expensive. To replicate the situations an autonomous vehicle might find itself in, you need to equip real cars with sensors, hire drivers, and send them out on the road. “It can take up to 10,000 hours of driving on the road to train software,” says engineer Pierre Olivier, chief technology officer at LeddarTech.
A world created from scratch
This is where the LeddarEcho software comes into play, developed by LeddarTech and the German company dSPACE to generate very large amounts of data at low cost. LeddarEcho uses digital twins to simulate an environment including roads, cones, pedestrians and, of course, self-driving cars. “It’s a bit like a big video game where vehicles drive around respecting the rules of the road”, illustrates Pierre Olivier.
LeddarEcho simulates and records the measurements that the sensors of the cars in the video game would have taken, and makes it a database that can be used to train artificial intelligence (AI). LeddarTech uses this data to develop its own products, and makes them available to some of its customers.
It’s a technology that greatly accelerates the deployment of perception solutions, because it allows you to start training the AI even before the [modèle de capteur] is available on the road.
Pierre Olivier, Chief Technology Officer at LeddarTech
“In real life, you can just take an hour of data for an hour of driving,” explains Mr. Olivier. “But in a simulation, you can speed up time, or run a hundred vehicles at the same time. Compared to reality, simulation also makes it easy to compare the algorithm to specific cases. For example, if we want to train him to recognize children getting out of a bus, it is difficult for a real driver to find himself faced with this specific situation hundreds of times in a row. With the software, all you have to do is program a scenario for this event to occur in a loop.
Obviously, these hyper-detailed simulations require extremely powerful computers, and therefore, inevitably, extremely expensive. “For synthetic data to be used, they must be very, very faithful to reality”, emphasizes Pierre Olivier. To be sure not to miss anything important, the algorithms still have to be confronted with a certain amount of real data, around 10% according to current standards.
A new type of sensor
The LeddarEcho software specifically makes it possible to simulate data produced by lidars, a type of sensor increasingly used in autonomous vehicles. The lidar essentially works like radar, sending out beams and measuring the time of return, but it uses light beams instead of radio waves. This allows it to be much more precise than a conventional radar. For the moment, radars are still mainly used in autonomous vehicles, but experts believe that lidars could take their place in the years to come, as their production costs drop.