At the main campus of the Korea Institute of Science and Technology (KIST) in Seoul's Seongbuk District, visited on the 16th, researchers were testing driver-tailored autonomous driving powered by a next-generation artificial intelligence chip. On a monitor running a self-driving simulation, a car circled roads near Sangam World Cup Stadium. A self-driving vehicle that had learned the owner's driving tendencies then retraced the route at nearly the same speed and along nearly the same path. What allowed it to reproduce the driver's individual habits was a neuromorphic chip linked to the autonomous driving system.
Kim Jae-wook, a senior researcher at KIST's Semiconductor Technology Research Division, said at the site that "neuromorphic technology in general implements the mechanism of the cerebellum, which handles the body's sense of balance and coordination." Kim added, "If a self-driving car also carries this chip, it can finely replicate each driver's individual driving skill on top of general autonomous functions such as responding to sudden situations."
Neuromorphic chips are next-generation AI semiconductors that imitate biological brain networks. Unlike conventional computing architectures, storage and computation are distributed across neurons and synapses and carried out together. Graphics processing units (GPUs), the best-known AI chips, by contrast use the so-called von Neumann architecture, in which the processor for computation and the memory for storage are separate. That means data moves back and forth repeatedly during operation and power consumption is excessive. As data travels between memory and processor, transfer speeds cannot keep up with the processor's fast computing performance.
Kim said that "existing deep-learning AI chips such as GPUs and neural processing units (NPUs) have to compute continuously in step with a calculation cycle, but a neuromorphic chip that borrows the brain's approach computes only when an event occurs." Kim added, "In energy terms it can be far more efficient than the existing approach of computing at every moment." Just as the brain generates signals only when stimulated, a neuromorphic chip exchanges electrical signals, or spikes, only when needed.
Concerns have been raised in particular that physical AI, which moves in the physical world, will be difficult to commercialize using the data centers that serve as AI infrastructure for large language models (LLMs). Running inference on a robot based on an AI model trained at a central data center requires enormous amounts of power. The ability to respond and adapt in real time as a device encounters its surroundings is therefore seen as essential.
Neuromorphic chips, with their strength in fine motion control, are also drawing attention as a new technology for commercializing humanoid robots. As part of a government-backed research and development program for next-generation intelligent semiconductors, KIST is developing core neuromorphic chip technology for real-time predictive control of physical AI systems. Humanoid robots on the market today struggle to maintain stable movement in unfamiliar settings such as slopes or sand. Robots carrying the cerebellum-mimicking neuromorphic chip that KIST is developing are expected to overcome those limits through the cerebellum's distinctive predictive mechanism.
Kim said that "when a person encounters terrain that differs from what was expected, they preemptively adjust their posture, and that is thanks to the cerebellum's predictive control function." Kim added, "Our goal is to link the chip to a humanoid in real time so that it can walk properly not only on flat ground but also in unstructured environments, including climbing slopes and changing direction."
The technology is also expected to be highly useful in wearable robots worn by rehabilitation patients and others. Kim's team, working with a team led by KIST researcher Lee Jong-won, applied a neuromorphic chip to a wearable robot and demonstrated muscle-strength assistance tailored to the wearer. The team plans to develop neuromorphic chip technology specialized for physical AI and secure intellectual property (IP) rights.
Kim said that "it is difficult to complete physical AI technology with GPUs or NPUs alone." Kim added, "Rather than fully replacing existing AI chips, neuromorphic chips will play a complementary role in commercializing physical AI." Kim also said, "In humanoids especially, there is attention on the potential complementary use of neuromorphic chips in areas that require massively parallel processing and real-time learning, like the functions of the cerebellum."
The United States, including Intel, leads in neuromorphic chip technology, which remains at an early stage with only some commercial products on the market. In South Korea, research institutes and academia are driving development. A team led by Kim Kyung-min, a professor of materials science and engineering at the Korea Advanced Institute of Science and Technology (KAIST), last month secured "neuromorphic neuron semiconductor technology" that uses noise occurring naturally in semiconductors as a resource for information processing. When the team converted body-activity signals and voice signals into spike signals to verify recognition performance, accuracy reached 94.8% for motion recognition and 95.0% for voice recognition. The technology mimics the brain by allowing a single piece of hardware to be reconfigured for signals of varying speeds and frequencies. It is expected to be used in signal processing for low-power neuromorphic systems.
