STORM: Hardware-Aware Tiny Transformer Co-Design for Low-Power Inertial Human Activity Recognition

Human Activity Recognition (HAR) from inertial sensors must run continuously on battery-powered wearables under tight latency, memory, and energy budgets. While tiny Transformers can be effective on inertial time series, end-to-end co-design across quantized inference and heterogeneous low-power platforms remains underexplored. We present STORM (Small Transformer for On-node Recognition of Motion), a deployment-oriented 19.7k-parameter 1D Transformer co-designed with X-HEEP, an open-source low-power single-core RISC-V SoC, and a tightly coupled streaming CGRA for nonlinear primitives (e.g., softmax). We build a cross-source 8-class benchmark by harmonizing 3 public datasets under a stringent, deployment-aligned protocol that exposes both cross-subject and cross-source shift. Using 1.280 s windows with 0.640 s stride, the protocol models continuous on-node HAR under cross-dataset generalization. After quantization-aware training and INT8 C inference export, STORM achieves 0.799/0.801 accuracy/macro-F1 on this benchmark. Deployed on an FPGA prototype of X-HEEP with the streaming CGRA backend, STORM requires 67.4 ms per inference at 100 MHz, while activity-based power analysis estimates a total inference energy of 632.4 μJ, satisfying the stride-driven real-time constraint. These results support the practical viability of compact attention-based HAR on low-power wearable-class embedded platforms.