Researchers at Oxford’s Department of Physics have demonstrated a new way to control a robot using brain-inspired computing hardware based on physical waves. Instead of performing computations using conventional hardware such as GPUs, the system harnesses the way waves naturally spread and interact to process information. The study, published in Nature Communications, represents a major step towards physical AI: integrating artificial intelligence into hardware that can operate in the physical world.
Many current AI systems run complex neural networks on powerful computer chips, often housed in large data centres or server farms. Training and operating these models can consume substantial amounts of energy, while sending information to and from remote data centres can slow responses and raise privacy and security concerns. These limitations are particularly important in robotics, where machines need to make decisions locally, efficiently and in real time while operating in unpredictable environments.
The experiment, performed by researchers from the Magnetism for Intelligent Devices (MIND) group at Oxford led by Dr Safeer Chenattukuzhiyil, drew inspiration from the collective dynamics of the brain, where complex behaviour emerges from populations of interacting neurons. A related framework, known as reservoir computing, uses the nonlinear dynamics of a physical system to transform information and requires training only at the output. Waves are particularly attractive for this approach because they naturally propagate, interact and interfere, generating complex patterns across space and time.
Starting with water waves in a metal container, the team showed that wave interactions can process sensory information and control a robotic vehicle in real-time. They then extended the same principle in simulations to nanoscale spin waves, pointing towards compact, energy-efficient neuromorphic hardware for future autonomous machines.
The researchers built a water-wave computing system using a circular metal container filled with water and electronically controlled actuators. The actuators dipped into the water at carefully controlled times, turning input data into waves whose interactions processed the information. An LED illuminated the water’s surface, while a camera captured the changing interference patterns projected onto a screen, making the physical computation directly visible.
The researchers first tested their wave-based hardware on robotic obstacle-recognition tasks. Information about the robot’s surroundings was converted into water waves, which travelled, interacted and formed distinctive interference patterns. After training, the system learned to recognise these patterns and identify different obstacle situations with up to 100% accuracy. The team then connected the system to a physical robotic car. The wave patterns determined how the car moved, enabling it to avoid obstacles autonomously. The researchers also demonstrated an event-driven approach in which computation was triggered only when the robot detected a change in its surroundings. This reduced unnecessary processing and offered a more energy-efficient and responsive way to control robots.
‘One of the most exciting aspects of this project is that we are bringing together two worlds: fundamental physics and autonomous robotics,’ comments Dr Safeer. ‘Instead of simply studying a physical phenomenon, here we harness the phenomenon itself to process information and make decisions for a robot.’
While the water system provides a visual demonstration of the technology, future AI chips will require ultrafast, solid-state wave systems operating at nanoscale dimensions and gigahertz speed. To explore this direction, the team also investigated spin waves, collective excitations of magnetisation that can propagate through magnetic materials. Micromagnetic simulations showed that a spin-wave reservoir only one micrometre in diameter could perform the same robotic recognition tasks. The results point towards electrically controlled, solid-state wave-based AI processors. These could be useful for the robotic applications demonstrated in the study and for a wider range of future AI applications.
Beyond the scientific and technological results, the work also shows how important scientific ideas can emerge from simple, unconventional experiments. At a time when frontier research is becoming increasingly expensive and resource-intensive, the team assembled the experimental platform largely from readily available items and components: a metal container, water, small motors, an LED, a camera and control electronics.
‘For us, this was a reminder that impactful research does not always have to begin with an expensive or complicated experimental system,’ said Professor Thorsten Hesjedal, a co-author of the study. ‘Sometimes, by looking differently at simple physical phenomena and cleverly building experiments using everyday objects around us, we can uncover entirely new possibilities.’
A patent application covering the underlying computing hardware has been filed with support from Oxford University Innovation (OUI), the University of Oxford’s innovation partner. Protecting the intellectual property is an important step on the journey from research to impact, creating a route for this novel approach to wave-based computing to be developed and applied beyond the laboratory. OUI is supporting the researchers to explore the IP’s market potential and the most effective pathways towards future applications.
Autonomous robotic operation controlled by wave-based neuromorphic hardware, J Zohar, D Pinna, G van der Laan, T Hesjedal, and CK Safeer, Nature Communications, September 2026