Practitioners in the field of TinyML lack so far a comprehensive, “batteries-included” toolkit to streamline continuous integration, continuous deployment and performance assessments of executing diverse machine learning models on various low-power IoT hardware. Addressing this gap, our paper introduces RIOT-ML, a versatile toolkit crafted to assist IoT designers and researchers in these tasks. To this end, we designed RIOT-ML based on an integration of an array of functionalities from a low-power embedded OS, a universal model transpiler and compiler, a toolkit for TinyML performance measurement, and a low-power over-the-air secure update framework—all of which usable on an open-access IoT testbed available to the community. Our open-source implementation of RIOT-ML and the initial experiments we report on showcase its utility in experimentally evaluating TinyML model performance across fleets of low-power IoT boards under test in the field, featuring a wide spectrum of heterogeneous microcontroller architectures and fleet network connectivity configurations. The existence of an open-source toolkit such as RIOT-ML is essential to expedite research combining artificial intelligence and IoT and to foster the full realization of edge computing’s potential.
