The Massachusetts Institute of Technology (MIT) is developing a self-driving car that can see through the ground to help it cope with bad weather.
Most self-driving cars use LIDAR but extreme weather such as snow and ice can affect the accuracy of sensors if they cannot see road features such as lane markings.
The team from the MIT Computer Science and Artificial Intelligence Lab (CSAIL) used “localizing ground penetrating radar” (LGPR) designed by the MIT Lincoln Laboratory and fitted it to a Toyota Prius.
The system was tested over 17km consisting of manual and autonomous driving in clear weather, rain and snow.
By mapping and localizing using features beneath the ground, the team, consisting of Teddy Ort, Igor Gilitschenski and Daniela Rus obtained features that are stable over time and maintain their appearance regardless of weather or lighting conditions.
They found that using LGPR provided precise localization for autonomous navigation without using cameras or LIDAR sensor.
The team admitted that their testing facility did not allow for high speed autonomous tests, a similar setup with a human driver at highway speed demonstrated successful localization.
The system also decreased in performance when the weather conditions during localization differed from when the maps were created, and it cannot perform global localization without prior use of GPS.
The team say future work will investigate modeling temperature and humidity induced reflection changes and global localization techniques for precise initialization even in the absence of prior GPS.
