Analog computing is experiencing a resurgence with hardware that processes and retains data in the same space, akin to biological neurons

A compact, lighter, and more energy-efficient computer, showcased at the University of Michigan, could aid in reducing weight and energy usage for autonomous drones and rovers, which has broader implications for self-driving vehicles.

The autonomous controller demonstrates among the lowest power demands recorded, according to the research published in Science Advances. It functions at just 12.5 microwatts—similar to a pacemaker’s consumption. In testing, a rolling robot equipped with the controller was able to chase a target moving in a zig-zag pattern down a corridor with comparable speed and precision to that of a traditional digital controller. In a second test, with a lever-arm that automatically adjusted its position, the innovative controller performed equally well.

“This research presents an innovative nanoelectronic device designed to transform hardware platforms capable of performing computations using neural network frameworks efficiently,” stated Xiaogan Liang, a professor of mechanical engineering at U-M and lead author of the research.

“These energy- and resource-efficient platforms are instrumental in promoting the miniaturization of robotic systems and vehicles.”

The high efficiency coupled with miniaturization is particularly crucial for applications such as drones and space rovers, where both weight and energy are of significant concern. Nevertheless, traditional autonomous vehicles could also take advantage of this technology. Research indicates that a billion hours of self-driving vehicle operation annually could use more power than all current data centers combined globally, according to previous studies.

Though analog computing has largely been overlooked in favor of digital systems for their reduced power usage and enhanced accuracy, it may now be an unexpected solution—thanks to a relatively novel circuit component.

The memristor, first introduced in 1971 and demonstrated in 2008, retains information based on its resistance to electrical currents. When powered by voltage, it diminishes the resistance presented to the next signal. Certain memristors can forget past signals over time and revert to their original resistance, mimicking relaxation observed in neurons. This is the variant that Liang’s team developed.

Since they function akin to neural networks, memristor networks can compute artificial neural networks far more effectively than traditional transistor-driven computers. Furthermore, retaining processing in the analog realm saves energy for sensors and actuators that are analog in nature, bypassing the energy costs associated with converting signals between analog and digital formats.

The team fabricated their memristor circuits in the Lurie Nanofabrication Facility at U-M by brushing a gold-tipped arm, about 30 microns (0.03 millimeters) thick, across a silicon chip—as one might rub a balloon on hair to create static electricity and make it adhere to a wall. This electrical charge directed vaporized bismuth selenide to form along eight intersecting lines roughly 15 nanometers (0.000015 millimeters) thick, arranged in a tic-tac-toe configuration. They subsequently plated electrodes of titanium and gold at the ends of each line.

Signals were transmitted through one electrode while being read at five electrodes on the opposite end of the chip, with each representing a neuron. In the study, video data from the rolling robot had to be converted into analog signals by a silicon processor before being processed through the memristor network. The silicon processor then transformed the output into control commands that allowed the robot to follow a red rectangular panel down a university corridor.

Analogously, for the lever arm, data regarding its position flowed into the memristor network through a silicon processor, producing the basis for operational instructions that directed the drone rotor to raise the arm to the proper height.

“Devices such as ours could allow robots to exhibit intuitive responses like humans do, akin to when you touch very hot water and reflexively pull your hand away. The control response may not be as precise, but it can be extremely swift,” remarked Mingze Chen, a recent Ph.D. graduate in mechanical engineering.

“Edge computing implies that information doesn’t need to be sent to a data center for processing, just as the nerves and muscles in our hand and arm can react without relaying information to our brains. Edge computing can be swifter, with reduced power consumption, since it eliminates the time and energy spent on data transmission.”

This research is funded by the National Science Foundation, and the device was developed at the Michigan Center for Materials Characterization.

Five of the authors in the study are undergraduate students enrolled in the Multidisciplinary Design Program at U-M.

The team has applied for patent protection with assistance from U-M Innovation Partnerships and is searching for partners to commercialize the technology.