Thinking Fast, Thinking Slow: Adaptive Multimodal Transformer-based Sensor Fusion for Depth Estimation on Ultra-low-power MCUs

Artificial intelligence (AI)-based multimodal sensor fusion is a relevant topic gaining ever more traction across ultra-low-power (ULP) embedded and cyber-physical systems, as it improves reliability, accuracy, and robustness under real-world constraints. However, adding more and more sensors to ultra-constrained sub-100 mW platforms requires balancing energy consumption against prediction accuracy. To achieve this ambitious goal, we present a novel adaptive AI methodology that combines multimodal sensor fusion (camera, ultrasound, and Time-of-Flight sensors) with a lightweight recurrent Transformer-based architecture (688 k parameters). We address the depth map estimation task with a mechanism that combines token propagation across iterations with incremental sensor utilization. At each iteration, a confidence-based gating mechanism dynamically decides whether to continue the computation by adding progressively richer but more power-demanding sensors as input. Token propagation ensures temporal consistency by forwarding context features across time. To deploy our algorithm and test a first real-world prototype, we design a novel printed circuit board featuring all three sensors, coupled with an ULP GWT GAP9 multicore System-on-Chip. When comparing our adaptive system against the same pipeline using all sensors and iterations on the NYUv2 dataset, we lose only 4.8% of the {\delta}1 accuracy in exchange for 90% energy saving (2.44 mJ/frame). Finally, our adaptive method marks only 5.6% lower {\delta}1 accuracy than MobileDepth despite using 9x fewer parameters. Compared with a state-of-the-art model also running on GAP9, our method improves {\delta}1 accuracy by 31.8% thanks to our adaptive sensor fusion while operating within the same average power budget (~400 mW).