Frank Markus
Technologue Look out: Super-sensor sees around corners, promises zero avoidable deaths. ’ve just learned it’s not entirely our fault, as Homo sapiens, that we cause 95 percent of car wrecks. That’s because we haven’t evolved the physiological ability to both see far enough ahead and process visual data sufficiently to eliminate “accidents.” I learned this when a California company called Neural Propulsion Systems (NPS) unveiled its AtomicSense Platform sensor fusion product, which it says can sense and interpret enough data to prevent crashes. NPS wrote a paper theorizing that perceiving the environment far enough, clear enough, and early enough to eliminate roadway deaths (that is, those from avoidable crashes) requires a data acquisition and processing rate of 100 terabits per second—that’s 10 million times the sensory data rate flowing from your eyes to your brain when you’re not drunk or distracted. So NPS devised a fully solid-state, system-on-chip (make that chips—lots of ’em) sensor suite to do just that, leveraging Department of Defense research. The AtomicSense system includes many cameras fused with three tech breakthroughs: digital multiband radar, multiple input/multiple output (MIMO) lidar, and atomic-norm computing. Today’s 77-GHz automotive radar is great at penetrating fog, but it paints a super-blurry picture of what’s out there, and raindrops can scatter it. However, radar at 1 GHz barely notices rain, the waves diffract to “see” around curves and corners, and they can even pass through some solid materials. A focused beam of highfrequency radar gives lidarlike resolution but is limited to line of sight. By employing four radar frequencies, NPS realizes radar’s full potential, achieving 100 times greater reliability than typical automotive radar units. Broadcasting digital pulses of radar virtually eliminates the risk of interference from other radar signals. NPS’ MIMO lidar uses a solid-state array consisting of multiple laser emitters and receivers, each focused on a small area and illuminated briefly, between 20 and 100 times per second. This flashing makes higherpower laser pulses safe for human eyes, which enables clear imaging at 540-plus yards—roughly twice the range of other lidars. Recent measurements at NASA’s Crows Landing runway detected a bicyclist at a recordsetting distance of 1,285 feet. The daunting task of fusing this rich radar and lidar data with camera imagery and making sense of it all is where atomic-norm computing comes in. This datacompression concept reduces measurements required to achieve the desired perception performance—it was originally developed to dramatically reduce the time patients had to spend in an MRI tube. AtomicSense
I
20 MOTORTREND.COM SEPTEMBER 2022
SOLID-STATE MIMO LIDAR
ATOMIC-NORM ALGORITHMS
SWAM RADAR
divides the scanned world into voxels—3-D pixels— measuring about 8x8x8 inches. Digitizing every bit of that space would generate 6.8 petabits/second of data to process, but in any scene, 99.0–99.9 percent of the voxels are empty and can be ignored. This knocks that total data stream workload down to a more manageable 100 terabits/second. By digitally steering the radar and lidar beams, the system can more thoroughly “interrogate” voxels of interest—those containing objects near the direction of travel. Empty voxels may get checked 20 times per second, while occupied ones are scrutinized 100 times. Atomic-norm computing allows the AtomicSense Platform to match the performance of conventional sensor suites with 1/50th the transmission power while yielding much higher resolution and vastly more hits on target. This allows it to reliably perceive pedestrians 650 yards away. The AtomicSense Platform sensor suite, anticipated for 2023 production, includes the software to identify objects all around the vehicle—including targets between parked cars or around corners that can’t be seen. It’ll be up to suppliers or automakers to integrate AtomicSense into an autonomous driving system, possibly increasing its capabilities by fusing data from high-definition maps, V2V or V2X communications, etc. Pricing (and form factor) will evolve to suit private vehicles, not just robotaxis and heavy trucks. The world has had a good look at “full self-driving” informed solely by cameras; I know I’ll feel 10 million times more comfortable with AtomicSense crunching 100 TB of triply redundant data Q.
Heavy trucks require greater braking distances, so long-range obstacle detection is critical to crash prevention. Shrinking both the form factor and price are necessary for broader adoption.