Single-Pass Hyperdimensional Computing on FPGA for Microsecond-Latency Adaptive Neuromodulation in Remote-Controlled Diving Cockroaches

Abstract

We propose a novel neural interface architecture for the remote-controlled diving cockroach biobotic platform that replaces conventional microcontroller-based pulse generation with a hyperdimensional computing (HDC) classifier implemented on a field-programmable gate array (FPGA). The system continuously processes two parallel data streams: native neural spike trains from the thoracic ganglia and antennal lobes and high-frequency bioimpedance spectroscopy data acquired across the electrode-tissue interface. A single-pass permutation encoding mechanism maps these 96-dimensional feature vectors into a 10,000-dimensional hyperdimensional space using pre-generated basis hypervectors and cyclic permutation operators implemented as barrel shifters on the FPGA. The classifier maintains eight class hypervectors corresponding to distinct stimulation parameter sets, and it selects the appropriate stimulation class by computing cosine similarity between the query hypervector and each class hypervector. A Hebbian-like online learning rule updates the selected class hypervector only when a reinforcement signal from the remote controller indicates successful motor response. The output pulse amplitude is further modulated by real-time impedance measurements to compensate for tissue compression during deep dives. The entire inference cycle, from feature acquisition to digital-to-analog converter update, completes in under one microsecond—a three-order-of-magnitude reduction compared to conventional software-based systems. The FPGA implementation on a Xilinx Artix-7 device consumes only 0.8 watts, well within the battery budget of the miniaturized backpack. This reformulated neural interface therefore enables real-time, adaptive neuromodulation for agile underwater maneuvers, addressing the critical latency bottleneck in existing biobotic control systems.