Embedding Approximate Nonlinear Model Predictive Control at Ultrahigh Speed and Extremely Low Power
Embedded systems require control algorithms that are safe and able to operate in embedded platforms with extreme limitations on energy, memory, and area footprint. Nonlinear model predictive control (NMPC) algorithms respect operational constraints to ensure safety but are typically challenging to implement on resource-constrained embedded systems at high speeds. This brief introduces a formalism for deploying an approximate NMPC control law on severely resource-constrained hardware by systematically leveraging approximate computing tools. The resulting field-programmable gate array (FPGA) implementation operates at extremely low power, is ultrafast, requires very small on-chip area, and consumes lower memory than cutting-edge implementations of embedded NMPC for systems of similar state-space dimension. Feasibility and stability guarantees are provided for the embedded controller by preemptively bounding the allowable approximation error in the hardware design phase. An FPGA-in-the-loop implementation exhibits speeds in nanosecond range with power consumption in <1 mW for 2-D and 3-D nonlinear systems.
Approximate computing, embedded systems, field-programmable gate array (FPGA), finite-precision, internetof- things (IoT), model predictive control (MPC), real-time systems.