New Chip Helps Robot Make Quick Decisions in Cat-and-Mouse Game
It takes a lot of energy to run with an ax, hammer, or bundle of explosives — and it’s even harder to do so when chasing a cunning mouse named Jerry. But while Tom and Jerry exist in a cartoon world, the need for energy efficiency transcends dimensions. Now, a new electronic chip successfully reduced power consumption and sped up decision-making for an autonomous mobile robot as it played a cat-and-mouse game reminiscent of the classic Tom and Jerry cartoons, but thankfully less violent.
Featured in Science Roboticsthe chip’s design is neuromorphic, or inspired by how people process external information to make decisions reflexively.
In real-life, autonomous mobile robots usually have limited amounts of power to use for both moving and making decisions. This constraint becomes a challenge during search and rescue missions as well as exploratory research.
“We [were] eager to design a neuromorphic chip to function as the brain of robots, allowing robots to have powerful intelligent processing capability and to cope with complex and unfamiliar environments,” Luping Shi, a professor in the Center for Brain-Inspired Computing Research at Tsinghua University and corresponding author of the study, wrote in an email.
Recently, interest in neuromorphic computing has grown because of its potential to optimize energy usage by neural networks and artificial intelligence (AI) algorithms.
“Due to the rapid development of artificial intelligence, the development of robots [has] entered a new era. However, computing hardware designed for robots is still lacking despite being highly desired,” wrote Shi.
The Secret to Multitasking
Robots are slowly entering the realm of real-time multitasking, but before they can truly cross that bridge, their computing hardware systems must be updated.
“The multitasking intelligent robot needs high computing power, high concurrency, low power consumption, high flexibility of resource scheduling, and easy-to-use computing hardware, so as to run intensive algorithms locally in real-time,” Shi and Songchen Ma, a Ph.D. candidate at Tsinghua University and leading author of the paper, wrote in an email.
Right now, processing units cannot efficiently meet AI’s needs for economical, concurrent, and adaptable processing. But neuromorphic systems, which approach computing more cooperatively, could be a solution.
To invent a new neuromorphic computing system that can holistically support neural networks, Ma and his colleagues first created a framework called Rivulet based on how different parts of the human brain work together to complete many tasks at once.
“It is challenging to design robot-friendly hardware, and there are few multitasking processors specifically designed for intelligent robots nowadays,” wrote Shi, Ma, and Weihao Zhang, a Ph.D. candidate at Tsinghua University and co-author of the research, in an email. “Inspired by the features of concurrency and cooperation in multiple brain regions, we propose [a] Rivulet execution model, which enables multiple neural network tasks to schedule resources on chip efficiently, flexibly, and simultaneously.”
using the Rivulet model, the authors built an electronic chip labeled TianjicX, as well as additional software. They then incorporated both into an autonomous mobile robot named Tianjicat and ran it through a demo cat and mouse game.
Gaming the System
during the game, Tianjicat had to track another non-autonomous robot’s movement, avoid obstacles, recognize objects, and identify sounds.
“The cat-and-mouse game is actually more complicated to implement than it seems. It needs the collaboration of multiple sensory information to deal with complex and dynamic scenarios, which is considered as the key for robots to achieve human-like intelligence,” wrote Shi, Ma, and Zhang.
Throughout the demo, the neuromorphic chip and its supporting software helped the robot accomplish intense multitasking. Further analysis revealed that TianjicX halved the amount of power Tianjicat consumed while processing information and significantly lowered delays between making decisions and acting on them. According to the researchers, these findings both show how their platform could be used in research on AI and in real-world scenarios.
“We provide an exploration platform for AI researchers and promote the study of brain-inspired intelligence,” wrote Shi and Ma. “Our platform also has great potential to be applied in edge computing scenarios due to its capability of multitask processing, such as autonomous driving, Internet of Things , [and] smart home [devices]†