Abstract

Model Predictive Control (MPC) is a state-of-the- art and robust control framework. However, its stringent per-formance requirements for computational infrastructure, and its need for real-time computation at the edge, hinder its widespread adoption in various application scenarios. This is particularly challenging for the MCUs commonly found on tiny robots. To address this computational challenge, we developed a flexible MPC solver hardware generation framework which includes a parameterized and programmable vector architecture template that accommodates instruction-level and data-level parallelism in vector and matrix functional units, and a model-specific fused architecture. Implementation of the proposed processor on the Ultra96 platform achieves up to a 9.73x speedup compared to existing generic solutions on MCUs. Moreover, end-to-end performance tests reveal that this speedup reduces the overall control error by 25.96%. Overall, the enhanced flexibility and performance of our proposed processor design open up the potential for MPC to be utilized in a broader range of applications.