MIT’s MapLite Tech Could Take Autonomous Vehicles Off-Road

Autonomous vehicles of the future could rely less on detailed maps and more on fine-tuned sensors to bring drivers down roads less traveled, thanks to researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL).

The lab team there has developed a navigation framework called MapLite, which gives up 3D maps in favor of a more minimalist system requiring only a rough roadmap like the one found in any consumer GPS system, and using its onboard sensors to then navigate the roads.

The system first sets a final destination and what researchers call the area “navigation goal” located within view of the car. Perception sensors then generate a path to get to that point, using Lidar to estimate the location of the road’s edges.

“We believe such a system has the potential to operate in a much wider area, and on a larger variety of roads, than those based exclusively on dense 3D maps,” CSAIL graduate student and lead author on a paper about MapLite, Teddy Ort, told The Connected Car.

Using a Toyota Prius outfitted with a range of Lidar and inertial measurement unit (IMUs) — sensors used for piloting drones — the research team autonomously drove on multiple unpaved country roads in Devens, Mass., and reliably detected the road more than 100 feet ahead.

Ort noted that while AV companies only test their fleets in major cities, where they’ve spent countless hours meticulously labeling a 3D urban landscape, if you live along the millions of miles of unpaved, unlit, unreliably marked US roads — you’re currently out of luck.

“Most autonomous navigation systems currently rely heavily on dense, 3D maps in order to navigate. These maps can be feasibly created for small urban areas,” Ort explained. “However, it is unlikely that such a solution would be scalable to all roads in the US due to the sheer enormity of the data that would have to be collected, stored, transmitted and maintained.”

Ort explained that while MapLite differs from other map-less driving approaches, which rely more on machine learning by training on data from one set of roads and then being tested on other ones, the framework still has some limitations.

For example, it isn’t yet reliable enough for mountain roads, since it doesn’t account for dramatic changes in elevation.

Ort said as a next step the team hopes to expand the variety of roads that the vehicle can handle — the ultimate aspiration being a system that performs as well as mapped systems but with a much wider range.

“I believe the biggest challenge we still face is figuring out exactly what mix of these sensors is the right one to balance the conflicting needs of reliability, range, accuracy and robustness to varying road conditions,” Ort said.

David Immerman, a 451 Research associate analyst for the Internet of Things, told The Connected Car that for automakers and other suppliers, exposing self-driving systems to different environments in a scalable way is proving problematic.

“We’ve seen an increase in traction in AV simulation software to help train these self-driving systems in the virtual realm while minimizing high AV testing costs,” he said.

On the sensor side, high sensor retrofitting costs for AV testing fleets is increasingly becoming a financial burden and time-intensive process, while the high traditional costs of Lidars alone have been a limiting factor for many in AV development.

“As of right now AVs are geared to function in an operational domain, whether that is to travel autonomously on highway or in a specifically mapped urban area,” Immerman said.

Ultimately, MapLite framework’s most promising application commercially could be for self-driving trucks, which make cross-country journeys in these remote environments where maintaining a real-time HD map may not be feasible.


Leave a comment

Your email address will not be published. Required fields are marked *