Car companies are working feverishly to improve the technology behind autonomous cars. But so far, even the most technologically advanced vehicles still struggle to navigate safely in rain and snow.
This is because these weather conditions wreak havoc on the most common detection methods, which usually involve lidar sensors or cameras. For example, in snow, cameras are no longer able to recognize lane markings and road signs, and the lasers of lidar sensors malfunction when, for example, objects fall from the sky.
MIT researchers recently wondered whether a completely different approach might work. What if we looked under the road instead?
Team from MIT Laboratory of Computer Science and Artificial Intelligence (CSAIL) has developed a up-to-date system that uses existing technology called ground penetrating radar (GPR) to send electromagnetic pulses underground that measure the specific combination of soil, rock and roots found in a given area. In particular, the CSAIL team used a particular form of GPR instrumentation developed in MIT Lincoln Laboratory called locating ground penetrating radar, or LGPR. The mapping process creates a unique fingerprint that the car can later exploit to locate itself when it returns to that particular plot of land.
“If you or I take a shovel and stick it into the ground, all we’ll see is a pile of dirt,” says CSAIL graduate student Teddy Ort, lead author of a up-to-date paper on the project that will be published in the Journal later this month. “But LGPR can quantify what is identified there and compare it to a map already created, so it knows exactly where it is without having to use cameras and lasers.”
During testing, the team found that in snowy conditions, the navigation system’s average margin of error was only about an inch compared to clear weather. Scientists were surprised to find that the rain had a bit more trouble, but still deviated by an average of only 5.5 inches. (This is because rain causes more water to soak into the ground, leading to a greater discrepancy between the originally mapped LGPR reading and the current soil condition).
The researchers said the system’s robustness was further supported by the fact that during six months of testing, they never had to unexpectedly step in to take over the wheel.
“Our work shows that this approach is actually a practical way to help autonomous cars navigate in difficult weather conditions without having to ‘see’ in the traditional sense using laser scanners or cameras,” says MIT professor Daniela Rus, director of CSAIL and senior author of a up-to-date paper that will also be presented at the International Conference on Robotics and Automation in Paris in May.
Although the team only tested the system at low speeds on a closed country road, Ort said existing work by Lincoln Laboratory suggests the system could easily be expanded to highways and other high-speed areas.
For the first time, autonomous system developers have used ground-penetrating radar, which has previously been used in areas such as construction planning, landmine detection, and even lunar exploration. This approach would not be able to operate completely on its own because it is unable to detect objects above the ground. However, its ability to locate in bad weather means it combines well with lidar and vision approaches.
“Before releasing autonomous vehicles on public streets, location and navigation must always be completely reliable,” says Roland Siegwart, professor of autonomous systems at ETH Zurich, who was not involved in the project. “The CSAIL team’s innovative and novel concept has the potential to bring autonomous vehicles much closer to real-world implementation.”
One of the main advantages of mapping an area using LGPR is that underground maps behave better over time than maps created using vision or lidar because the features of above-ground maps are much more susceptible to change. LGPR maps also take up only about 80 percent of the space taken up by the customary 2D sensor maps that many companies exploit in their cars.
While the system represents an significant advance, Ort notes that it is not yet ready for road exploit. Future work will need to focus on designing mapping techniques that can combine LGPR datasets to account for multi-lane roads and intersections. Additionally, current equipment is bulky and 6 feet wide, so significant advances in design must be made before it is diminutive and lithe enough to fit in commercial vehicles.
Ort and Rus wrote the paper with CSAIL intern Igor Gilitschensky. The project was supported in part by the MIT Lincoln Laboratory.