Self-driving cars that depend on light-based image sensors typically have trouble seeing in harsh conditions like fog. Now, researchers at MIT have invented a new sub-terahertz-radiation receiving system that could help these autonomous vehicles navigate if traditional methods fail.
Sub-terahertz wavelengths are wavelengths between microwave and infrared radiation on the electromagnetic spectrum. They can easily be seen through fog and dust clouds whereas the typical infrared-based LiDAR imaging system that self-driving cars currently use cannot. To identify an object, a sub-terahertz imaging system sends an initial signal via a transmitter then a receiver measures the absorption and reflection of the rebounding sub-terahertz wavelengths. This sends a signal to a processor, which recreates an image of the object in question.
However, the challenge is integrating these sub-terahertz sensors into autonomous vehicles. There needs to be a strong output baseband signal from the receiver to the processor for accurate object-recognition. Typical systems, composed of discrete components that generate these signals, are big and costly and the smaller, on-chip sensor arrays create signals that are too weak.
The IEEE Journal of Solid-State Circuits published an online paper on Feb. 8 by the researchers. The paper details a two-dimensional, sub-terahertz receiving array on a chip that can better pick up and interpret sub-terahertz wavelengths even if there is a lot of signal noise.
For this to work, they used a scheme of independent signal-mixing pixels known as “heterodyne detectors.” This is typically quite hard to densely implement into chips. The researchers dramatically reduced the size of the heterodyne detectors so that a lot of them could fit into one chip. They created a condensed multipurpose part that can down-mix input signals, synchronize the pixel array, and generate strong output baseband signals all at the same time.
The researchers created a prototype that had a 32-pixel array meshed onto a 1.2-square-millimetre device. These pixels are about 4,300 times more sensitive than the pixels in today’s best on-chip sub-terahertz array sensors. With further work, the chip could be used in self-driving cars and autonomous robots.
Ruonan Han, a co-author of this paper and an associate professor of electrical engineering and computer science, and the director of the Terahertz Integrated Electronics Group in the MIT Microsystems Technology Laboratories (MTL) explained that one of the biggest inspirations for this project is to have improved ‘electric eyes’ for self-driving vehicles and drones. Han added that their cost-effective, on-chip sub-terahertz sensors would have a useful role in LiDAR when the environment is harsh.
Han also worked on the paper with first author Zhi Hu and co-author Cheng Wang, both PhD students in the Department of Electrical Engineering and Computer Science and a part of Han’s research group.
Decentralisation is what made this design possible. In this way, a single pixel — known as “heterodyne” pixel — produces the frequency beat (the difference in frequency between two incoming sub-terahertz signals) and the “local oscillation,” an electrical signal that modifies the frequency of an input frequency. This “down-mixing” process generates a signal in the megahertz range that a baseband processor can be easily read.
In much the same way that LiDAR calculates the time it takes a laser to hit an object and bounce back, the output signal can be used to determine how distant an object is. What’s more, mixing the output signals of many different pixels, and steering the pixels in a particular direction, can allow for high-resolution images of a landscape. This enables both the detection and the recognition of objects, which is vital for self-driving cars and autonomous robots.
Heterodyne pixel arrays function only when the local oscillation signals from all pixels are synchronized. This means that a signal-synchronising method is required. For centralized designs, there is a single hub that shares local oscillation signals with all pixels.
Typically, receivers of lower frequencies employ these designs but they can cause problems at sub-terahertz frequency bands, where producing a high-power signal from a single hub is quite challenging. As the array scales up, the power shared by each pixel diminishes, lowering the output baseband signal strength, which depends significantly on the power of the local oscillation signal. In turn, a signal produced by each pixel is not very strong, causing low sensitivity. Some on-chip sensors have begun using this design, but they only have eight pixels.
Thus, the new decentralized design that the researchers’ came up with address this scale-sensitivity trade-off. Each pixel produces its own local oscillation signal, which is used for receiving and down-mixing the incoming signal. Furthermore, an integrated coupler synchronizes its local oscillation signal with that of its neighbour. This gives more output power to each pixel because the local oscillation signal does not come from a global hub.
Han explained that it is easy to think of the new decentralized design as an irrigation system. A typical irrigation system has one pump directing a powerful stream of water through a network of pipelines, which then distribute the water to many sprinklers. Each sprinkler’s water flow is much less intense than the initial flow from the pump. If you want the sprinklers all to have the same strength, you would need another control system.
However, by using the new decentralized design, each site gets its own water pump, getting rid of the need for connecting pipelines, and each sprinkler gets its own output of powerful water. Moreover, the pulse rates are synchronized because each “sprinkler” can communicate with its neighbour.
Thus, Han explained that with this design there’s basically no ceiling for scalability. One can have as many sites as one wants, and each section will pump out the same amount of water, and all the pumps will pulse in unison.
However, this new decentralized design has the potential to make the footprint of each pixel much bigger. This hinders the large-scale, high-density integration in an array fashion. For their model, the researchers mixed different functions from four usually separate parts: the antenna, downmixer, oscillator, and coupler, into a single “multitasking” piece given to each pixel. This enables a decentralized design of 32 pixels.
For the system to ascertain the distance of an object, the frequency of the local oscillation signal needs to be stable.
Therefore, the researchers made sure that their chip had a component known as a phase-locked loop, which locks the sub-terahertz frequency of all 32 local oscillation signals to a stable, low-frequency reference. Since the pixels are coupled, their local oscillation signals all share identical, high-stability phase and frequency. This guarantees that one can derive meaningful information from the output baseband signals. This entire concept reduces the loss of signals and increases control.