Energy-efficient computer and robotics technologies inspired by nature

Since the AI revolution, data centres have been guzzling electricity. Because the digital chips in our computers are not suited to artificial intelligence, CogniGron, a multidisciplinary research centre at the University of Groningen, has opted for neuromorphic technology: a way of working inspired by the human brain and nervous system.

TEXT: TIMO KÖNNEN

Globally, data centres use around 1.5 per cent of all electricity. That still sounds modest. But consumption is growing steadily and that growth appears to be accelerating: from 2020 to 2024, consumption rose by 12 per cent per year, and last year by 17 per cent. This is mainly due to the AI revolution: consumption by data centres built specifically for AI grew by 50 per cent in 2025. This is not only a problem because ofCO₂ emissions and climate change; it also exacerbates issues on the electricity grid, leading to power shortages.

This rapid growth is driven by the electricity-hungry ‘deep neural networks’ powering the AI of OpenAI, Google, Meta and other tech companies. These networks are large computer programmes that function somewhat like the brain. However, they run on servers which, with their silicon chips packed full of transistors, bear no resemblance to a brain. The computers perform complex calculations to simulate brain behaviour to some extent. ‘Neural networks nowadays have millions of connections and are constantly growing,’ says Alejandro Pequeño-Zurro, a researcher in neuromorphic sensing within Elisabetta Chicca’s bio-inspired circuits and systems group at the University of Groningen. ‘A server consumes about one kilowatt when interpreting camera images – as much as a kettle.’ This stands in stark contrast to our brains, which need only a few watts to process the visual information from our eyes.

Processors and memory

Why do computers waste so much energy? A major culprit is the separation between the processor and memory, which means that sending data back and forth requires a great deal of electricity.

The human brain makes no distinction between processor and memory. The circuits in the brain – systems of interconnected nerve cells (neurons) – do not learn by storing data in a large storehouse, but through constant self-evaluation and adaptation. Each circuit acts as both a mini-processor and a mini-memory (see box: Neural domino).

Pequeño-Zurro is therefore studying neuromorphic computing, with the aim of using the principles of the brain to design better computers and systems. The University of Groningen brings together multidisciplinary research in this field at the Cognitive Systems and Materials Centre (CogniGron), established in 2018. The aim is to create a computer chip that is ten thousand times more efficient than today’s computers.

To encourage cross-fertilisation, the research projects take various approaches, ranging from materials for artificial neurons to complete devices that are controlled by conventional chips but still operate in a ‘natural’ way. Within this diverse research, a number of recurring themes can be identified.

Bottom-up construction

‘The fact that AI systems have become so large and energy-hungry is down to the top-down approach taken by the tech giants,’ says Pequeño-Zurro. ‘First, the required performance is determined; then a network is designed to meet those criteria. Higher performance mainly means a larger network with greater resource usage.’ Nature, on the other hand, has always worked from the bottom up. It was only after extensive experimentation with single-celled organisms that organisms consisting of multiple cells emerged. In the course of further evolution, ‘proven’ systems were also used as building blocks for larger organisms.

To apply this approach to AI, researchers at CogniGron-
re developing small circuits that will later serve as building blocks for larger systems. Pequeño-Zurro: ‘We want to develop basic elements, inspired by the brain’s fundamental neural networks.’ The idea, he explains, is that this will eventually enable the construction of powerful computer systems that are nevertheless small and energy-efficient. Pequeño-Zurro points to a small rectangle measuring a few square centimetres, mounted on a board which is itself connected to a PC via cables. On this chip, he explains, there are circuits running machine-learning algorithms inspired by the brain. The PC is currently testing how well these circuits interact with their environment.

The ‘neurons’ of these circuits are electronic circuits, comprising components such as capacitors, resistors and signal amplifiers. It is a form of analogue electronics: unlike the digital variant, which can only operate with 0 (no) and 1 (yes), the neural elements on this chip can handle all values in between. Just like the human brain, this chip learns because the connections between the neurons become stronger the more often a signal passes through them.

Natural hardware

These types of chips are produced by specialist companies, which manufacture them in series or to the customer’s specifications. Their capacity is still small, often just a few hundred neurons. The number of circuits on digital chips doubles every two years (Moore’s Law), but as there is not yet a mass market for analogue chips, development is proceeding much more slowly. Every step in miniaturisation simply costs a great deal of money. However, there are also materials that conduct electricity better the more current passes through them: memristors (memory resistors). These can replace circuits on a neuromorphic chip that strengthen the connection between two neurons the more frequently it is used. The fundamental materials research required for memristors is being carried out at CogniGron by, amongst others, Professor Tamalika Banerjee. She developed a memristor consisting of thin layers of metal-oxygen compounds and founded the company IMChip to mass-produce these memristors for neuromorphic chips. The German firm Siemens intends to use the technology to develop AI for military drones.

Minimisation

Minimalism is perhaps nature’s most striking characteristic: everything happens with a minimum of materials and energy. For example, a desert ant has no more than a large nerve cluster for a ‘brain’, yet with this it manages to find its way back to the nest every day after walking hundreds of metres. This fascinated Professor of Robotics Ming Cao. He is working on autonomous robots for domestic and industrial use, capable of carrying out tasks independently. In a video shown by Cao, a dog-like robot searches for an object in a space that is new to it. It does so in a rather human way – not in a rigid manner, but purposefully. This places high demands on a deep neural network, says Cao. ‘And it guzzles energy, which is of course incompatible with the aim of an autonomous robot for every household. Ultimately, therefore, we will have to switch to neuromorphic systems.’

Cao felt that the desert ant’s navigation provided a good example from which to learn. Biologists discovered that the ant has a built-in pedometer, which also tracks changes in direction. Cao based his computer model for a pulsating neural network (see box: Neural domino) on this, and it proved capable of navigating well in simulated desert environments. The trick was to vary the speed of the connections between the neurons, just as in nature. This gives the neurons greater computational capabilities. Such a system could be of interest for small drones and robots that need to operate in places where there is no GPS. ‘To test the model in the real world, we’re going to build a small robot controlled by a chip containing the network.’

Data pre-processing

Picking up objects is an underestimated human skill. Robots struggle with it, unless they are dealing with a fixed set of objects in a consistently identical environment. The fact that we rarely let anything slip or crush it to pieces is down to the rapid feedback loop between our tactile sensations and our movements. That speed is partly due to the fact that our senses are more than just passive conduits. Pequeño-Zurro: ‘The tactile sensors in the hand are linked in clusters. The signals within each cluster are first compared; only then is a message sent to the brain.’

This pre-processing – computing at the edge – means fewer nerve pathways are required and the process is faster: minimalism at its best. Other senses work in the same way; for example, the human eye itself ‘sees’ boundaries between areas of colour. The neuromorphic researchers would like to apply the same principle to robots and other devices equipped with sensors. Using a computational model of a robotic hand tasked with determining the orientation of a row of rods, Pequeño-Zurro is working on proving that pre-processing does indeed result in a gain in speed. ‘This kind of fundamental research deepens our understanding of the building blocks we work with. And it provides us with useful insights into biological systems.’

Mimicking the sense of touch

You’re unlikely to see a robot feeling an object to determine what it’s made of. ‘The sense of touch is fundamental for humans and animals,’ says Pequeño-Zurro, who specialises in the sense of touch in robots. ‘Both for handling objects and for social interaction. But it is a neglected area of research in robotics, where vision dominates. This sometimes leads to peculiar situations, such as an AI system trying to determine what something is made of using camera images.’

With neuromorphic computing, artificial touch can now also be developed. In Pequeño-Zurro’s laboratory, there is a device with a drum around which a strip of flexible material – fabric, paper, carpet – can be wound. An artificial finger rests on the surface of the material. As the drum rotates, the bumps or ridges on the material cause sensors in the finger to vibrate. The signals from the sensors are sent to a neural network. This consists of various circuits, each of which can resonate at a different frequency (see box: Neural domino). Almost always, several of these resonances occur simultaneously. It is this combination that allows the material to be recognised. This is probably also how the brain works: the time interval between the stimuli provides essential information for the sense of touch. ‘The circuits in the brain responsible for registering touch are very precise in terms of timing. The intensity of the touch, however, is perceived much less precisely,’ says Pequeño-Zurro. Cogni-Gron is now working with other partners on a demonstrator – a device for identifying materials in an industrial setting. Pequeño-Zurro: ‘In this way, we want to show that the Dutch knowledge ecosystem is capable of utilising neuromorphic techniques.’

Strength in numbers

Finally, nature also offers a collective form of minimalism: small animals that work together in swarms, colonies or shoals. ‘Termites, despite their tiny size, are collectively capable of building nests metres high,’ says Bahar Haghighat, a robotics researcher at the Faculty of Science and Engineering. She draws inspiration from such swarm behaviour and designs systems of small, mobile or flying robots that can, for example, clean the windows of a block of flats or monitor air quality in a city. ‘The little robots work quickly as a collective; they are cheap, easy to replace and pose no danger to people.’ They can come in handy in unexpected places. A commonly used method for checking the condition of a large machine or structure, such as a bridge, is to detect and analyse vibrations. Sensors are attached to the surface for this purpose. The companies offering this method analyse the measurement data in the cloud. That is not a comforting thought for all owners. ‘An operator of, for example, a power station or a bridge does not want to run the risk of the data falling into the hands of malicious parties,’ says Haghighat.

Her solution: to have the inspection carried out by a swarm of mobile robots, or ‘rovers’. ‘They travel in random directions across the structure, stopping regularly to listen for vibrations. In this way, they locate defects such as cracks and holes.’ The rovers do not need the cloud, as they process their measurement data themselves. They do this in a self-learning manner: a neural network runs on their microcontroller.

Of course, the rovers cannot perform calculations as complex as a server in the cloud, but they compensate for this by covering a much larger proportion of the surface area as a group – dozens or more of them. Haghighat wants to further improve performance using neuromorphic techniques. ‘To get the most out of such small devices with limited battery capacity, it makes sense to use neuromorphic chips. That’s what I’m working on now.’