Automotive manufacturers are racing to enhance the sophisticated technologies powering autonomous vehicles. Despite these advancements, even the most cutting-edge self-driving cars struggle to operate safely when encountering rain and snow-covered roads.
These challenging weather conditions significantly disrupt the most prevalent sensing methodologies, typically relying on lidar sensors or camera systems. During snowy conditions, cameras fail to identify lane markings and traffic signs, while lidar sensor lasers malfunction when precipitation interferes with their operation.
Researchers at MIT have recently explored whether a fundamentally different approach might solve these weather-related challenges. Specifically, they investigated the potential of looking beneath the road surface rather than above it.
A team from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) has engineered an innovative system that leverages existing ground-penetrating radar (GPR) technology. This system sends electromagnetic pulses beneath the road surface to measure the unique combination of soil composition, rocks, and root systems. The CSAIL team specifically utilized a specialized form of GPR instrumentation developed at MIT Lincoln Laboratory known as localizing ground-penetrating radar, or LGPR. This mapping process generates a distinctive underground fingerprint that autonomous vehicles can subsequently use to precisely localize themselves when returning to that specific terrain.
"If humans were to dig into the ground with a shovel, we would only perceive dirt," explains Teddy Ort, a CSAIL PhD student and lead author of a new paper about the project, scheduled for publication in the IEEE Robotics and Automation Letters journal later this month. "However, LGPR can quantify the specific elements beneath the surface and compare them to previously created maps, enabling the vehicle to determine its exact position without relying on cameras or lasers."
During testing, the team discovered that in snowy conditions, the navigation system's average margin of error was merely about one inch compared to clear weather operation. The researchers were surprised to find that the system experienced slightly more difficulty in rainy conditions, though it still maintained an average error of only 5.5 inches. (This occurs because rainwater absorption into the soil creates greater disparity between original LGPR map readings and current soil conditions.)
The system's reliability was further validated over six months of testing, during which researchers never needed to unexpectedly intervene and take manual control of the vehicle.
"Our research demonstrates that this approach offers a practical solution for helping autonomous vehicles navigate adverse weather conditions without requiring traditional 'sight' capabilities through laser scanners or cameras," states MIT Professor Daniela Rus, director of CSAIL and senior author on the paper, which will also be presented in May at the International Conference on Robotics and Automation in Paris.
Although the team has only tested the system at low speeds on a closed country road, Ort notes that existing research from Lincoln Laboratory suggests the technology could readily be adapted for highways and other high-speed environments.
This marks the first instance where ground-penetrating radar has been employed in autonomous vehicle development, though the technology has previously been utilized in construction planning, landmine detection, and even lunar exploration. The approach cannot function entirely independently, as it cannot detect above-ground obstacles. However, its ability to provide precise localization in adverse weather makes it an excellent complement to lidar and vision-based systems.
"Before deploying autonomous vehicles on public roads, localization and navigation systems must demonstrate complete reliability under all conditions," remarks Roland Siegwart, a professor of autonomous systems at ETH Zurich who was not involved in the project. "The CSAIL team's innovative concept has significant potential to bring autonomous vehicles much closer to real-world implementation."
One significant advantage of mapping areas with LGPR is that underground maps maintain their accuracy longer than maps created using vision or lidar technologies, since above-ground features are much more likely to change over time. Additionally, LGPR maps occupy only about 80% of the storage space required by traditional 2D sensor maps that many companies currently employ for their vehicles.
While the system represents a crucial advancement, Ort acknowledges that it remains far from being road-ready. Future research will need to focus on developing mapping techniques that enable LGPR datasets to be integrated for handling multi-lane roads and intersections. Furthermore, the current hardware is bulky and six feet wide, necessitating significant design improvements before it can be made compact and lightweight enough for integration into commercial vehicles.
Ort and Rus co-authored the paper with CSAIL postdoc Igor Gilitschenski. The project received partial support from MIT Lincoln Laboratory.
