Ultra-low-power chips help make small robots more capable

A brain-inspired ultra-low-power hybrid chip can help palm-sized robots collaborate and learn from their experience. Combined with a new generation of low-power motors and sensors, the new application-specific integrated circuits (ASICs)-running at milliwatts-can help smart swarm robots run for hours instead of minutes.

To save power, the chip uses a hybrid digital-analog time-domain processor, where the signal’s pulse width encodes information. Neural network ICs are suitable for model-based programming and collaborative reinforcement learning, and may provide greater reconnaissance, search and rescue, and other mission capabilities for small robots.

Researchers at Georgia Institute of Technology demonstrated robotic cars powered by unique ASICs at the 2019 IEEE International Solid-State Circuits Conference (ISSCC). The research was sponsored by the Defense Advanced Research Projects Agency (DARPA) and Semiconductor Research Corporation (SRC) through the Brain-Inspired Computing Power Autonomous Intelligence Center (CBRIC).

“We are working to provide intelligence for these very small robots so that they can understand their environment and move autonomously without infrastructure,” said Arijit Raychowdhury, associate professor of electrical and computer engineering at Georgia Tech. One goal, we want to bring the concept of low-power circuits to these very small devices so that they can make their own decisions. There is a huge demand for very small but powerful robots that require infrastructure. “

The cars demonstrated by Raychowdhury and graduate students Ningyuan Cao, Muya Chang and Anupam Golder passed through a rubber-covered arena surrounded by cardboard block walls. When they search for targets, robots must avoid traffic cones and each other, learn in the environment, and constantly communicate with each other.

Cars use inertial and ultrasonic sensors to determine their location and detect objects around them. Information from the sensors is passed to a hybrid ASIC, which acts as the “brain” of the vehicle. Then the instructions point to the Raspberry Pi controller, which sends instructions to the motor.

In a palm-sized robot, three main systems consume power: motors and controllers for driving and manipulating the wheels, processors and sensing systems. In cars made by the Raychowdhury team, low-power ASICs mean that motors consume most of their power. “We have been able to reduce computing power to a level where the budget is dominated by motor demand,” he said.

The team is working with motor partners using micro-electromechanical (MEMS) technology, which can operate at lower power than traditional motors.

“We want to build a system where the sensing power, communications and computer power and drive are at about the same level, about hundreds of milliwatts,” said Raychowdhury, an ON Semiconductor associate professor at ON Semiconductor. Electrical and computer engineering. “If we can use efficient motors and controllers to make these palm-sized robots, we should be able to provide hours of running time on a few AA batteries. We now know what kind of computing platform we need to provide, but We still need other components to catch up. “

In time-domain calculations, information is transmitted on two different voltages, which are encoded in pulse widths. This provides the circuit with the energy efficiency advantages of analog circuits and the robustness of digital equipment.

“The chip size has been cut in half, and power consumption is one-third that of traditional digital chips,” Raychowdhury said. “We use multiple technologies in our logic and memory designs to reduce power consumption to the milliwatt range while meeting target performance.”

Because each pulse width represents a different value, the system is slower than digital or analog devices, but Raychowdhury says that speed is sufficient for small robots. (Milliwatts are 1 / 10,000th of a watt).

“For these control systems, we don’t need circuits that work at several gigahertz because the equipment is not moving fast,” he said. “We are sacrificing a bit of performance for extremely high power efficiency. Even if the computer is operating at 10 or 100 MHz, this is enough for our target application.”.

65nm CMOS chip is suitable for two learning methods suitable for robots. The system can be programmed to follow model-based algorithms and can learn from its environment using an augmentation system that promotes better and better performance over time-just like a child who learns to walk by colliding things .

Raychowdhury said: “You start the system with a predetermined set of weights in the neural network so that the robot can start in a good place without crashing or providing incorrect information immediately.” The environment will have some structure that it will recognize and some that the system must learn. Then the system will make its own decisions and will evaluate the effectiveness of each decision to optimize its actions. “

Communication between robots allows them to collaborate to find a target.

“In a collaborative environment, a robot needs to understand not only what it is doing, but also what other people in the same group are doing,” he said. “They will work to maximize the total rewards of the group, not the individual rewards.”

As their ISSCC demonstration provides a proof of concept, the team is continuing to optimize the design and is developing a system-on-chip to integrate computing and control circuits.

“We want to implement more and more functions in these small robots,” Raychowdhury added. “We have shown what is possible, and what we do now needs to be strengthened by other innovations.”

Solemnly declare: The copyright of this article belongs to the original author. The reprinted article is only for the purpose of disseminating more information. If the author’s information is marked incorrectly, please contact us as soon as possible to modify or delete it. Thank you.


Related Posts

Leave a Reply

Your email address will not be published. Required fields are marked *