Across the spectrum of imaging sensors fused into the autonomous vehicle driving system, radio detection and ranging (RADAR) is arguably best positioned. It’s the most reliable sensor, that works in any light conditions, weather conditions, and reaches 300 meters in sensing by Azimuth. It’s also the only sensor that can measure the doppler velocity naturally and the price tag is much cheaper than the parallel LiDAR solutions. The only problem is that the existing RADARs in the market are not sufficient for today’s Level 2 of autonomous driving, or even adaptive cruise control. Arbe is the first mover to develop and demonstrate an ultra-high resolution 4D imaging radar with post-processing and simultaneous localization and mapping (SLAM) that’s bridges the gap between a radar and optics for all levels of vehicle autonomy.
Arbe was started by Kobi Marenko, Dr. Noam Arkind, and Oz Fixman after working and founding at Taptica (sold to Marimedia for $14 million). Arkind, a PhD in Mathematics and Computer Science from the Weizmann, worked for Israel’s Aerospace Institute as a control algorithm specialist for spacecraft motion and landing, and eventually joined Taptica as a Machine Learning specialist. Marenko is a serial AI entrepreneur on his second AI venture, having started Taptica, and another startup previously called Logia Group (acquired by Digital Turbine), and the COO and CFO of local news station Channel 10. The team originally converged to work on a swarm intelligence based solution for the drone market, but quickly pivoted to the automotive angle.
From the beginning, they envisioned signal processing and AI on top of layer for an off-the-shelf Radio Frequency (RF) chipset and Digital Signal Processor (DSP). But poor performance results from the market’s RF chipsets led them to build their own. They brought onboard an expert team in chipsets, for Radio Frequency Integrated Circuit (RFIC) and cracked the RF problem but concluded their signal processing and software cannot operate on a general processor. So they designed a digital processor and Application-Specific Integrated Circuit (ASIC).
Their chipset has 12 receiving channels and 24 transmitting channels. The performance of the chip of Arbe is better than the basic parameters of the competitors who have a four by three array; which mean 12 virtual channels, whereas Arbe is processing over 2,300 channels. Every frame they process at 4G is about 30 GB of data and “we couldn’t buy a processing chip in the market that can process the large amount of data that’s generated imaging RADAR”, which is why Arbe opted to design their own processor, according to Marenko.
Arbe’s patented system provides full mapping in four dimensions (distance, azimuth, elevation, and Doppler speed – the radial velocity of the point or object), and high resolution and long range (300 meters), at an accessible price. The system facilitates autonomous driving at all levels, from driver support (ADAS), to fully autonomous, driverless travel.
“The cycle of building hardware is long; a year to design, a year for fab production in testing and qualifying, and the length in total is in years, while the barriers to entry is also equal to years” explained Marenko.
Arbe’s proprietary, Artificial Intelligence based, post-processing software stack includes a radar-based Simultaneous Localization and Mapping (SLAM) solution powered by AI. The SLAM algorithms perform real-time clustering, tracking, and self-localization as well as false-target filtering and radar-based and radar-camera based object classification. It is important to note that performing SLAM while utilizing radar data is a patent pending innovation.
Their AI algorithms needs to understand if the detected object is a human being and not a tree and it needs to make the calculation where it will be in a second, since this has a major influence on what the car needs to do. It also needs to eliminate false targets. In the last phase, it also fuses with the camera, and the other sensors in the suite to classify and match the objects from the multiple sensors equipped on the AV.
“We use an array of algorithms, starting from deterministic and prior based, all the way to deep learning and proprietary, sophisticated models” explained Marenko.
“Today, because the radar resolution is so low, you’re not doing any post processing that’s related to real fusion” explained Marenko. Because Arbe is the first company in the market to build a radar that generates a real picture, there’s suddenly an opportunity to run recognition systems to try to understand what is happening in this scene. “It’s really like the first algorithms of computer vision 20 years ago.” The radar is integrating into the digital age, and because they’re the first company to have that picture, they’re similar to “the first company that was able to codify video” compared Marenko. “And with this new high resolution image generated from 4D imaging, we are also able to analyze it.” Arbe’s RADAR generates a point cloud picture for analysis.
The startup is using AI and SLAM algorithms to identify and localize objects and velocity, and its surroundings according to the doppler signature. It’s compacted into a modular form and is suited for up to Level 5 autonomy and any vehicle size.
The main advantage of radar, aside from the fact that it costs 10% of LiDAR systems, is that it can reach a better range and penetrate anywhere in any conditions according to Marenko. They believe their RADAR will solve a majority of the problems and eliminate redundancies in the sensor architecture.
The startup delivered products to over 10 OEMs and new mobility players, dedicated to drive testing for autonomous systems.
The startup’s breakthrough super RFIC and patented signal processing sets them aside from the pack. Their signal processing for post processing of the 4D Imaging data detects, separate and tracks moving objects in real time. They have 10 provisional patents and have raised a total $23 million of funding. They have offices in Tel Aviv, Beijing, and Silicon Valley, and they’re growing.
The startup works with OEMs for premium cars that want to implement Level 2 and 3 of autonomous driving, and the new mobility players trying to implement robo taxi services at level 4 & 5 autonomy. For consumer cars, Arbe expects their chipsets to be integrated within the next three years, providing Level 3 autonomous driving experience. At this pace, the era of autonomy in cars is soon becoming a reality.
