Autonomous vehicles can follow general road rules, recognizing road signs and road markings, noting pedestrian crossings and other well-known traffic control features. But what to do outside of well-marked roads, traveled far and wide? On many roads outside the cities, the paint was faded, the signs were overgrown with ivy and trees, unusual crossroads appeared that were not marked on the maps.
What should the next autonomous car do when the rules are unclear or missing? What should its passengers do when they discover that their car cannot take them where they are going?
Most of the challenges in developing advanced technologies include handling rare or unusual situations or events that require performance that goes beyond the normal capabilities of the system. This definitely works in the case
of autonomous cars. Some road examples may include navigating through repair areas, meeting a horse or buggy, or meeting with graffiti that resembles a brake light. Outside the road, opportunities include absolutely all manifestations of the natural world, such as trees blocking the road, floods and large puddles – or even animals blocking the path.
At the Center for Advanced Automotive Systems at the University of Mississippi, scientists have taken on the task of teaching algorithms to respond to circumstances that are almost never encountered, difficult to predict, and difficult to recreate. They tried to place autonomous cars in the most difficult scenario: they drove the car to an area that he had not seen and did not know before, without any reliable infrastructure, such as road paint and road signs, in an unknown environment, where cactus and white can be encountered with equal probability. bear
In the process, they combined the technology of the virtual and real worlds. They created advanced simulations of realistic scenes in the open air, with the help of which they trained artificial intelligence algorithms to read the flow from the camera and classify what they saw: trees, sky, open paths, possible obstacles. Then they transferred these algorithms to a specially created test all-wheel drive car and sent it to a dedicated test site, where they then checked the performance of the algorithms that collect data.
Let’s start with the virtual
Engineers developed a simulator capable of creating a wide range of realistic outdoor scenes through which vehicles could move. The system generates a variety of landscapes with different climates, forests and deserts, shows how plants, shrubs and trees grow over time. It can also simulate changes in weather, sunlight and moonlight, as well as the exact position of 9000 stars.
In addition, the system simulates the readings of sensors commonly used in autonomous vehicles, such as lidars and cameras. These virtual sensors collect data, which is then fed to neural networks as valuable data for training.
Let’s build a test track
Simulations are only as good as they reflect the real world. The University of Mississippi has acquired 50 acres of land on which scientists are developing a test track for self-driving off-road vehicles. The plot is great – there are slopes at an angle of 60 degrees and a lot of different plants.
Engineers have identified some of the natural features of this land, which, they expect, will be particularly difficult to cope with self-driving cars, and reproduced them exactly on the simulator. This allowed them to directly compare the simulation results with the actual attempts to navigate this earth. Ultimately, they will create similar real and virtual pairs of other types of landscapes to improve the capabilities of the cars.
Collect additional data
A test vehicle was also created – Halo Project – with an electric motor and sensors with computers that can navigate through a variety of off-road environments. The Halo Project car is equipped with additional sensors to collect detailed data about its real environment; they help build virtual environments to run new tests.
Two lidar sensors, for example, are mounted at cross-angles on the front of the car, so that their rays scan the approaching ground. Together they can provide information on how rough or smooth the surface is, as well as read data on grass and other plants and objects on the road.
In general, research scientists gave some interesting results. For example, they showed promising hints that machine learning algorithms that train in simulated environments can be useful in the real world. As is the case with most studies on autonomous transport, there is still a long way to go. Perhaps, together, they will help make self-driving vehicles not only more functional on modern roads, but also the more popular and common method of transportation.