Artificial intelligence is often associated with heavy computing demands and high energy consumption—obstacles that limit its use in the Internet of Things (IoT), where sensors and devices typically run on minimal power and processing capabilities. However, researchers from the E-MINDS project have developed methods to make AI efficient enough to run on extremely limited hardware, paving the way for smarter, low-power applications in industry and beyond.
A joint initiative involving the COMET K1 center Pro2Future, Graz University of Technology (TU Graz), and the University of St. Gallen, the E-MINDS project has demonstrated how specialized AI models can operate locally on devices with just 4 kilobytes of memory. These models are able to perform tasks such as identifying sources of interference in ultra-wideband (UWB) localization systems, without relying on cloud computing or external processors.
Instead of running large, general-purpose models, the team focused on building highly specific ones tailored to individual tasks like distance estimation. To make even these small models efficient enough, researchers applied a series of optimization techniques. The outcome is a modular system in which several specialized AI models are used instead of one large model. For example, separate models handle interference from metal walls, people, or shelving in localization tasks.
An orchestration model on the device identifies the type of interference and loads the appropriate model from a server in under 100 milliseconds—a speed suitable for real-time industrial uses such as automated warehouses.
Another technique, known as Subspace Configurable Networks (SCNs), allows models to adapt to input data rather than requiring a different model for every scenario. These SCNs proved effective in image recognition tasks, such as object classification, by offering improved processing speed and energy efficiency on IoT hardware. In one test, these models processed images up to 7.8 times faster than comparable systems relying on external servers.
To further reduce computational load, the researchers applied methods such as model folding, quantization, and pruning. Folding compresses the model’s mathematical structure without drastically impacting accuracy. Quantization involves converting floating-point numbers into simpler integers to reduce processing effort. Pruning trims unnecessary parts of a model post-training, keeping only the components critical to its core function.
Each of these methods contributed to balancing performance with minimal resource use—ensuring models remain accurate while fitting into the constraints of embedded systems.
Although the primary focus of the project was on improving UWB localization for precise positioning of drones, shuttles, or robots in industrial environments, the researchers see a much broader potential for their techniques. One example is enhancing the security of keyless car entry systems by verifying the proximity of a key rather than just its signal. Other possibilities include improving battery life in smart home devices or enabling libraries to track books more effectively.
Beyond model miniaturization, the project also explored how to efficiently transfer AI models onto tiny devices, ensuring deployment is as lightweight as the models themselves.
The E-MINDS project has laid a technological foundation that can benefit a wide range of industries. By bringing advanced AI capabilities to the smallest devices, the team has opened up new possibilities for real-time, on-device intelligence across sectors where energy, space, and hardware constraints have long been limiting factors.
By Impact Lab
