On the radar revolution, deadlines and access to the fourth dimension
In my previous articles about autonomous trams and locomotives I have already mentioned radar technologies. Radar is widely used in the automobile industry for standard active and passive safety features. Solutions for highly automated control systems (including autonomous vehicles) require more flexible and advanced technologies. At Cognitive Pilot, we have a dedicated radar team. Until late 2019, this team worked as a design house, developing custom solutions for automobile manufacturers and parts vendors. We’re currently moving to a new business model and preparing a line of mass-produced radar systems for a wide range of customers — from DIY projects to startups and test fleets. The solutions used in Cognitive Pilot projects will serve as a basis for ready-to-use consumer products that can be divided into three categories: MiniRadar, Industrial and 4D Imaging. Such devices are actively used in a variety of industries, so they deserve a more detailed overview.
Access to the fourth dimension
Even though car radar units are traditionally referred to as 3D, they can’t determine the height of an object. To an outsider, this may seem like a marketing trick. The physical properties of the signal (the Doppler effect) enable such units to measure three parameters [R, Az, V]: distance and angle (azimuth) to the object, as well as velocity and its sign. The typical set of sensors for a self-driving car includes video cameras, long-range all-weather radar units in active safety systems and a lidar unit capable of accurate 3D scene measurements. The lidars are expensive (for example, Uber installs units that cost around $120,000) even though their only purpose is to obtain a 3D cloud of points. Additionally, they make it impossible to forgo other sensors.
We started considering the development of a radar unit that could replace the costly lidar. Omitting the intermediary stages of analysis, calculations and assessments, I’ll just say that we found it possible. As early as summer 2017, we created the first operational proof of concept (POC) with an external waveguide antenna array. We had to use precision equipment to manufacture the antennae for the required frequency band (up to 77 GHz), so this setup was too bulky and expensive for mass production, but it worked for a POC. Furthermore, the initial hardware components we used were imperfect and included multiple analog solutions. The design didn’t feature any moving parts and was based on digital beamforming array architecture — not unlike the one used in fighter jets. Most importantly, the POC demonstrated the conceptual possibility of product implementation.
Then, as CES 2018 was approaching, we decided to develop the world’s first industrial version of a 4D radar unit with a planar antenna array (read on to find more about it) capable of measuring range, azimuth, elevation angle and velocity [R, Az, Ev and V]. To complete the project on time, we had to completely rework the VHF part on a tight schedule. We were facing an issue with our partners, who took a month-and-a-half to manufacture a board for our project from a special VHF material, and we needed several iterations. For our commercial prototype (which was also a POC in terms of the printed circuit board material), we contracted with a like-minded domestic partner — AO NIIPP (Research Institute of Semiconductor Devices) in Tomsk. They took about a month to complete all the iterations of the antenna implementation at a low-temperature co-fired ceramics (LTCC) production line. I am particularly grateful to Mr. Evgeny Monastyrev for this work.
As a result, we obtained a hair-thin ceramic plate with a large area, which we used for a planar antenna. The next step was mounting it inside the radar casing by attaching it to a titanium base (the coefficient of thermal expansion of titanium matches that of ceramics and prevents the board from cracking at temperature fluctuations). The timing was tight, so we had to bring it from Moscow by plane in our luggage. Then we had to assemble the radar unit, test it and prepare a demo video before January 4.
The bearing capacity of a ceramic plate is low, so we had to glue it to a hard base. NIIPP performed this operation, too, with the use of a special pressing machine. The process took a dramatic turn on December 27–28 when the only antenna snapped as we were assembling the device. Our colleagues in Tomsk showed incredible understanding and supported us in our quest for fame overseas: they launched their production line on December 30 and 31 to give us the assembled component by January 1. It took us two days to mount, configure, and fine-tune the hardware, and by January 4, we had made a demo video to show our device in action. Later on, we switched back to imported material with the required frequency properties, but back then, at the end of 2017, only a domestic vendor could manufacture the prototype we needed on time.
Radar design and working principles
We wanted to develop relatively cheap and compact devices without moving parts, affordable even to small startups and DIY enthusiasts. Since there is no workaround for the laws of physics, developing the VHF part was a major challenge: to obtain a high angular resolution, we needed a full-fledged phased array antenna. All our radar units are equipped with planar (microstrip) antenna arrays implemented as lines of a special shape on boards. Due to high (up to 81 GHz) frequencies, the conventional composite epoxy materials used in consumer electronics weren’t an option. We needed a material with a low level of signal attenuation per linear inch.
We had another problem related to the electronics within the device. We had to save space while ensuring high capacity. Radar units process received data onboard; they don’t just transmit an analog signal. As a result, the user gets coordinates of objects as well as the direction and speed of their movement. The last couple of decades have seen rapid advancements in microelectronics, and now the market offers highly integrated systems capable of realizing many useful features. The latest-generation models make it possible to create a single-chip radar unit, even though the device would be relatively simplistic. The chip features an analog part that includes receiver and transmitter units, an AD converter and hardware-based accelerators that perform Fourier transformation among other functions. The digital unit includes a DSP (digital signal processor) and an ARM processor. The level of data processing is aligned with the capabilities of the sensor: radar units with a small number of channels and a minimal angular resolution are equipped with chips that match their needs.
All Cognitive Pilot radar sensors are based on the MIMO (Multiple Input Multiple Output) principle, which is a coded pattern method for multiplying the channel bandwidth. Receiver and transmitter units are geometrically dispersed and can emit signals by turns (temporal channel distribution), as different code chains (code channel distribution) or by combining the two approaches. This is a way of enhancing a radar’s characteristics without making its design more complex and expensive. The main advantage is the smaller number of necessary receiving channels. Our smallest radars, for instance, have three transmitters and four receivers. The transmitters emit different code chains simultaneously, similarly to what happens in 3G and CDMA. Four physical receivers pick up these chains separately and assemble the signal of each transmitter. The resulting 12 virtual receiving channels increase the resolution threefold without any modifications to the hardware setup. We also could have achieved such a result by using eight more receiving lines and additional AD converters, but this would have made the device more complex, driving up the costs substantially.
We implement the entire development stack in-house: the VHF part, the electronics and other hardware components and its visual aspect. The hardware is a very important part of a radar unit, but there is more to it. The performance of a radar unit and the data you can obtain from it depend on algorithms, such as object detection, secondary processing filters, and code chains. So, we develop all of the above too — all the algorithms of the mathematic model starting from signal generation are created in-house. To that end, the single-chip solution that we used in our MiniRadar units features sophisticated microcode software with a range of subsystems, for instance, for the control of analog peripheral components or hardware-based accelerators. The solution offers flexibility in configuration, making it possible to optimize data flows and their transmission from one unit to another.
Product line
MiniRadar units are ready-to-use single-board solutions that can be connected through CAN or SPI port (depending on the modification) to the on-board computer of a car or even to an Arduino microcontroller, which is popular among DIY enthusiasts. Other series are similar in terms of antenna arrays (with a horizontal viewing angle of 120°–150°) but feature a more complex design and consist of several modules (VHF, digital processing, power supply, and interfaces). They have many more channels, which means a much higher angular resolution: for example, the Industrial model features 32 receivers backed by serious computing capacity. The primary analog-digital board with a set of transceivers and an antenna array required strengthening with additional digital processing units (boards) with a high-performance DSP and an Ethernet adapter with power supplied over a network cable.
With a horizontal viewing angle of 120°–150°, an Imaging 4D radar unit also emits a vertical beam. Knowing the time of reception and disappearance of the reflected signal, one could take a bearing, obtain the vertical angle of the beam directed at the object and determine the third coordinate of the point. The 4D radar version that reached the stage of mass production was licensed for use by some of our clients. Since then, we have made progress and now are preparing a new solution with more advanced technologies than those we used in 2017. These new technologies will be free from any contractual limitations and will be available to a wide audience.
Devices from different series vary in functionality and the quality of yielded results. The MiniRadar series was intended for such car features as the emergency braking system, adaptive cruise control or blind spot control. Industrial series sensors can be used in automated industrial suites, monitoring systems and locomotives, while the advanced Imaging 4D series was designed for self-driving transport.
Future plans
Since early 2020, we’ve been trying to make Cognitive Pilot radar technologies available to customers en masse. We have plenty of technologies to offer, including a synthesized aperture for ultrahigh-resolution images, object signature assessment based on micro-Doppler interference, super-resolution and localization based on radar data.
Our solutions vary in technological complexity and price so that users can find the best fit for their projects. We have many plans for the future and even more tasks to handle at the moment (our R&D team is never bored). Stay tuned to learn more about the technologies we use.
