(A) Schematic sketch for illustrating the implementation of the memristive network–based RC system for rover control through processing time-sequential sensory signals. Here, voltage-based analog sensory signals carrying spatiotemporal information components are input to the memristive reservoir. These input signals are differentiated and nonlinearly mapped to a high-dimensional data space based on the temporal contexts of the input sensory signals and are quantitatively represented by the reservoir state vector X(t), which is constructed from the voltage readings at multiple neuron terminals. Afterward, the state vector is multiplied by a pretrained weight matrix W(t) to export the output signals Y(t) for controlling the testing rover. (B to D) Training data acquisition for emulating PID control of a robot rover for performing target-tracking navigation: (B) snapshot captured from the training video, showing the PID-controlled rover tracing after a red-moving target (the inset view is a snapshot from the ESP32-based internet-of-things (IoT) camera on the rover); (C) exemplary target coordinate data plotted as the function of time points; (D) exemplary motor signal data generated by a digital PID controller, plotted as the function of time points.
(A) Schematic sketch for illustrating the implementation of the memristive network–based RC system for rover control through processing time-sequential sensory signals. Here, voltage-based analog sensory signals carrying spatiotemporal information components are input to the memristive reservoir. These input signals are differentiated and nonlinearly mapped to a high-dimensional data space based on the temporal contexts of the input sensory signals and are quantitatively represented by the reservoir state vector X(t), which is constructed from the voltage readings at multiple neuron terminals. Afterward, the state vector is multiplied by a pretrained weight matrix W(t) to export the output signals Y(t) for controlling the testing rover. (B to D) Training data acquisition for emulating PID control of a robot rover for performing target-tracking navigation: (B) snapshot captured from the training video, showing the PID-controlled rover tracing after a red-moving target (the inset view is a snapshot from the ESP32-based internet-of-things (IoT) camera on the rover); (C) exemplary target coordinate data plotted as the function of time points; (D) exemplary motor signal data generated by a digital PID controller, plotted as the function of time points.
(A) Schematic sketch for illustrating the implementation of the memristive network–based RC system for rover control through processing time-sequential sensory signals. Here, voltage-based analog sensory signals carrying spatiotemporal information components are input to the memristive reservoir. These input signals are differentiated and nonlinearly mapped to a high-dimensional data space based on the temporal contexts of the input sensory signals and are quantitatively represented by the reservoir state vector X(t), which is constructed from the voltage readings at multiple neuron terminals. Afterward, the state vector is multiplied by a pretrained weight matrix W(t) to export the output signals Y(t) for controlling the testing rover. (B to D) Training data acquisition for emulating PID control of a robot rover for performing target-tracking navigation: (B) snapshot captured from the training video, showing the PID-controlled rover tracing after a red-moving target (the inset view is a snapshot from the ESP32-based internet-of-things (IoT) camera on the rover); (C) exemplary target coordinate data plotted as the function of time points; (D) exemplary motor signal data generated by a digital PID controller, plotted as the function of time points.
A smaller, lighter and more energy efficient computer, demonstrated at the University of Michigan, could help save weight and power for autonomous drones and rovers, with implications for autonomous vehicles more broadly.
The autonomous controller has among the lowest power requirements reported, according to the study published in Science Advances. It operates at a mere 12.5 microwatts—in the ballpark of a pacemaker. In their testing, a rolling robot using the controller was able to pursue a target zig-zagging down a hallway with the same speed and accuracy as with a conventional digital controller. In a second trial, with a lever-arm that automatically repositioned itself, the new controller did just as well.
“This work introduces a groundbreaking nanoelectronic device designed to revolutionize hardware platforms that can efficiently compute with neural network architectures,” said Xiaogan Liang, U-M professor of mechanical engineering and corresponding author of the study.
“These energy- and resource-efficient platforms pave the way for advancing the miniaturization of robotic systems and vehicles.”
The high efficiency and miniaturization is especially important for applications like drones and space rovers, in which both weight and energy are at a premium. However, conventional autonomous vehicles could also benefit from the technology. A billion hours of autonomous vehicle drive time per year could consume more power than today’s data centers combined worldwide, according to prior research.
Analog computing, all but abandoned for digital’s lower power consumption and higher precision, may seem an unlikely hero—but a relatively new circuit element is changing the game.
The memristor, proposed in 1971 and first demonstrated in 2008, stores information in its resistance to electrical currents. When it is exposed to a voltage, it reduces the amount of resistance it will impose on the next signal. Some memristors can forget previous signals over time and return to their original resistance, a behavior that is similar to relaxation in neurons. This is the type that Liang’s team built.
Because they already function a lot like neural networks, memristor networks compute artificial neural networks much more efficiently than conventional transistor-based computers do. In addition, for sensors and actuators that are analog themselves, keeping processing analog saves the energy costs of converting signals between analog and digital.
The team built their memristor circuits in the Lurie Nanofabrication Facility at U-M by rubbing a gold-tipped arm, roughly 30 microns (0.03 millimeters) in diameter, across a silicon chip—like rubbing a balloon on your hair so that it will stick to a wall with static electricity. The electrical charges then guided vaporized bismuth selenide to accumulate along eight criss-crossing lines about 15 nanometers (0.000015 millimeters) thick, arranged similar to a tic-tac-toe board. They then plated on electrodes of titanium and gold at the ends of each line.
They injected signals through one electrode and read them out at five electrodes on the other side of the chip, each representing a neuron. In the study, camera data from the rolling robot had to be converted to analog signals in a silicon processor before running through the memristor network. The silicon processor then converted the output into control instructions that enabled the robot to follow a red rectangular panel down a university hallway.
Similarly, for the lever arm, data about the position of the arm went into the memristor network through a silicon processor, and it produced the foundations of instructions for running the attached drone rotor to lift the arm to the correct position.
“Devices like ours could enable robots to have intuitive behaviors like human beings, the way you might touch very hot water and pull your hand back. The control response may be less accurate, but it can be very fast,” said Mingze Chen, a recent Ph.D. graduate in mechanical engineering.
“Edge computing means the information doesn’t have to travel to a data center for processing, like the nerves and muscles in our hand and arm can react without sending the information to our brains. Edge computing can be faster, with lower power consumption, because we don’t spend time and energy on transmitting data.”
The research is funded by the National Science Foundation. The device was studied at the Michigan Center for Materials Characterization.
Five of the study authors are undergraduate students participating in the Multidisciplinary Design Program at U-M.
The team has applied for patent protection with the assistance of U-M Innovation Partnerships and is seeking partners to bring the technology to market.
Did this article help?
If you found this reporting useful, please consider supporting our work with a small donation. Your contribution lets us continue to bring you accurate, thought-provoking science and medical news that you can trust. Independent reporting takes time, effort, and resources, and your support makes it possible for us to keep exploring the stories that matter to you. Thank you so much!