Most of the existing autonomous-vehicle (AV) localization strategies try to replicate how human drivers perceive the world. But reliable autonomy requires new approaches, argues Tarik Bolat, the chief executive of WaveSense. “Technologists need to widen the aperture of what data is available to the vehicle and not just replicate how humans drive,” he said. Bolat’s solution is to use stepped-frequency, continuous-wave ground-penetrating radar to peer into a place where human eyes can’t, up to 10 feet (3 m) into the ground.
WaveSense and its competitor, Geophysical Survey Systems, Inc. (GSSI) both use technology developed by MIT’s Lincoln Laboratory and deployed in 2013 to help American troops in Afghanistan avoid landmines. Both companies modified the MIT technology to enhance its performance for automotive applications. WaveSense and GSSI similarly apply their technologies to localization rather than perception. It requires strapping the localizing ground-penetrating radar (LGPR) to the undercarriage of a vehicle and creating a map of the underground soil density, roots, rocks, cavities and utility infrastructure.
When vehicles similarly equipped with LGPR subsequently travel on the same path, they use the ground’s signature (something like a fingerprint) to precisely locate position relative to objects. Because the location markers are below ground, the technology is not affected by severe weather. It also works in places where the built environment lacks differentiated visual infrastructure, or when GPS drops out for brief periods.
A tale of two radar companies
The two companies – the only firms actively pursuing ground-penetrating radar – license the same fundamental technology from MIT’s Lincoln Laboratory. However, WaveSense makes the stronger claim on the intellectual rights, claiming to be the “worldwide exclusive licensee of a broad patent portfolio owned by MIT for localizing ground-penetrating radar.” The Somerville, Mass.-based company, which was formed in 2017, is a startup dedicated to automotive applications.
Meanwhile, GSSI is a 50-year old company based in Nashua, N.H., utilizing the LPGR for a wide range of industries, including construction, geophysics and archeology. The two firms similarly use a relatively small set of low-resolution pixelated images, indexed against other location technologies such as GPS, to pinpoint a vehicle’s location down to a few centimeters. They claim a similar degree of precision, about four centimeters.
WaveSense says the main output of a comparison of the mapped versus tracked underground is a precise longitude and latitude. GSSI claims something similar but claims the primary deliverable to the vehicle’s system is the degree to which the mapped environment correlates to the tracked environment. “If you tell me where you are within 50 meters,” said David Cist, GSSI’s chief technology officer, “in most cases, we will give you an accurate location within four centimeters.” GSSI is working with a single major German automaker that is testing for autonomous-vehicle localization.
WaveSense is collaborating with a handful of automakers and Tier-1 suppliers, primarily focusing on the United States and Europe. It’s also speaking with Asian companies. WaveSense’s Bolat exudes confidence in the technology. “WaveSense is uniquely solving gaps that are unsolved today by the standard [AV compute] ‘stack’ and don’t have an alternate line of sight on being solved,” Bolat said.
But GSSI’s Cist is more cautious. “LGPR technology still needs to be proved out on many levels,” he said. “It’s very easy to overstate the current position of the technology. I think the technology has a long way to go before it would ever be in a commercial vehicle.”
Thin bridges and parking decks
The two companies espouse the same belief in the technology regarding its ability to make up for the shortcomings of other localization strategies that can be stymied by a snowstorm – or tricked by a change in the environment – a recently fallen tree, for example. They both bolt a box under the vehicle with radar antennae using a frequency of between 100 and 400 MHz rather than the 77 GHz of a typical front-facing radar.
Both companies argue that you don’t need much data, or powerful AI, to do the necessary frame matching. WaveSense says its rate typically is about 100 scans per second. In contrast, GSSI uses a variable speed of about 10 per second, possibly faster in the future. GSSI matches frames in five dimensions. There are other significant points of divergence between companies. The hardware used by WaveSense is a box measuring about one foot by two feet by one inch. Bolat says the encasement is rugged enough to withstand the punishment of the road. But he declined to comment on the product’s design.
Cist said GSSI relies on LGPR hardware with larger dimensions: approximately five feet by two feet by one-third of a foot, spanning the entire width of the vehicle. “There are 12 antennas laterally aligned to cover as much road as possible,” he explained. Cist also said there needs to be enough air to provide back reflection and thereby allow the antennas to do their job.
Bolat from WaveSense says automakers working with his company have specific applications in mind, such as automated parking in a garage, or lane-centering. A parking garage might present column after column that are evenly spaced. An open highway in the middle of the country could have lane markings that are “chewed up a little bit.” Those above-ground roadway features can be challenging for vision-based systems. But Bolat expresses no doubt that WaveSense’s ground-penetrating solution can solve for those problems – or even the challenge of steel bridges.
“Dirt roads are not an issue, typically,” he said. “And we were just working with a major OEM on autonomous parking looking at 20 centimeters thick of a concrete parking deck. But you’re able to generate enough character to do sub-five-centimeter-level localization accuracy in that environment.”
GSSI’s Cist said he believes that some roadways have soil variability that “only needs the first meter or so to localize accurately.” But the system’s algorithm can adjust for seasonal changes in the ground moisture or monotonous soils. “Some soil conditions or road and bridge structures compromise accurate localization altogether, which is why complementary methods must be used when LGPR fails,” he said.
He is even less sanguine about LGPR’s ability to handle parking structures. “You ought to take self-parking with an enormous grain of salt,” he warned. “Our radar penetrating three meters, what’s it going to with a parking garage that’s half-a-meter thick? It’s not going to give you any unique information because the decks are just rows and rows of rebar.”
These distinctions in attitude notwithstanding, both companies could have bigger business problems to solve. Dan Smith, the chief executive of Capstone Financial Group, a San Jose-based investment bank focusing on auto technologies, is impressed with the technology but has questions about its practicality. Smith said he believes that attaching a new type of device to the underbelly of a vehicle, and the testing and validation required to ensure that it’s safe and free of liability issues, is onerous. “I think the idea of ground-penetrating radar is pure genius,” he said. “But commercializing it is going to be tough.”Continue reading »