Researchers from Graz University of Technology (TU Graz), Pro2Future, and the University of St. Gallen have developed methods to enable artificial intelligence (AI) to operate on low-resource Internet of Things (IoT) devices. The project, conducted under the E-MINDS initiative, aimed to allow AI models to function locally on small devices with limited memory and processing capacity, eliminating the need for external computing resources.
The research demonstrated that AI models could run on ultra-wideband localisation devices with only 4 kilobytes of memory. These models can identify interference sources affecting positioning accuracy, such as metal structures or human movement. The system employs a modular approach, using several small, task-specific models instead of a single general model. An orchestration model on the device determines the type of interference and activates the corresponding AI model within approximately 100 milliseconds, a speed deemed sufficient for industrial settings like warehouses.
Among the techniques applied were subspace configurable networks (SCNs), which adapt to varying input data, enabling faster and more energy-efficient image recognition tasks. Quantisation and pruning were also employed to simplify models by converting floating-point numbers to integers and removing non-essential model components, respectively. These methods aimed to maintain acceptable accuracy while reducing model size and energy consumption.
The project focused on wireless ultra-wideband localisation for industrial automation applications, including tracking drones and robots in environments with interference. However, the findings are considered applicable to other domains, such as enhancing the security of keyless car entry systems and extending the battery life of smart home devices.
According to the project team, the collaboration combined expertise in embedded systems, machine learning, and localisation technologies. The work is positioned as a foundation for future AI-driven applications on constrained hardware platforms.
Photo credit: Lunghammer – TU Graz
