Brain-like computers turn to rolling robots, with 0.25% of the power required by traditional controllers

A smaller, lighter, more energy-efficient computer presented by the University of Michigan can help save weight and power from autonomous drones and wanderers, with a wider impact on autonomous vehicles.

According to research published in Science Advances, autonomous driving controllers report the lowest power requirements. It runs in the form of 12.5 microwatts, on the pacemaker’s court. In their tests, the rolling robot using the controller was able to pursue a goal of twists and turns with the same speed and accuracy as a traditional digital controller. In the second trial, the new controller did a great job with the lever arm that was automatically repositioned.

“This work introduces a groundbreaking nanoelectronics designed to revolutionize the hardware platform that can be efficiently computed through neural network architectures,” said Xiaogan Liang, a professor of mechanical engineering and study author.

“These effective platforms for energy and resources pave the way for advancing the miniaturization of robotic systems and vehicles.”

High efficiency and miniaturization are especially important for applications such as drones and space wanderers, where both weight and energy are premium. However, traditional self-driving cars can also benefit from the technology. According to previous research, there are a billion hours of self-driving cars driving every year that consumes more power than today’s data centers in total.

Analog computing almost abandons digital lower power consumption and higher accuracy, and seems unlikely to be hero, but relatively newer circuit elements are changing the game.

The memorandum was proposed in 1971 and was first proven in 2008 for its resistant storage to currents. When exposed to voltage, it reduces the amount of resistance it will apply on the next signal. Some memos will forget the previous signals over time and restore their original resistance, a behavior similar to relaxation in neurons. This is the type of team building of Liang.

Since they already work like neural networks, Memristor Networks compute artificial neural networks more efficiently than conventional transistor-based computers. Furthermore, for sensors and actuators of the simulation itself, keeping the simulation process can save energy costs of converting signals between analog and digital.

The team built their remote circuits in UM’s Lurie nanofabrication facility by rubbing a golden arm (about 30 microns (0.03 mm in diameter) of silicon chip, crossing the silicon chip, such as rubbing a balloon on the hair, in order to stick it to the wall and stick it to the wall with static electricity. The charge-guided fibrous selenization of the evaporated fibrous selenization, stacking about 15 nanometers (0.000015 mm) thick along eight crisscrossing lines, similar to the TIC-TAC-TOE plates. They then plated the titanium and gold electrodes at the ends of each row onto the titanium and gold electrodes.

They inject signals into the signal through one electrode and read them in five electrodes on the other side of the chip, each representing neurons. In the study, before running through the Memristor network, the camera data of the rolling robot must be converted into analog signals in the silicon processor. The silicon processor then converts the output into control instructions, allowing the robot to follow the red rectangular panels of the university corridor.

Similarly, for the lever arm, data about arm position enters the Memristor network through the silicon processor and generates a command basis for running the connected drone rotor to lift the arm to the correct position.

“A device like ours can enable robots to have intuitive behavior like humans, where you might run into very hot water and pull the hands back. The control response may not be very accurate, but it can be very fast,” said Mingze Chen, a recent Ph.D. Graduated from mechanical engineering.

“Edge computing means that information doesn’t have to go to a data center for processing, such as the nerves and muscles in our hands and arms that can react without sending the information to our brain. Edge computing can be faster and less power consumption because we don’t spend time and effort transferring data.”

The research was funded by the National Science Foundation. The device was studied at the Michigan Materials Characterization Center.

Five study authors are undergraduate students participating in a multidisciplinary design program at UM.

The team has applied for patent protection with the assistance of UM Innovation Partnership and is seeking partners to bring technology to market.