Offloading Reservoir Computing-Based Hand-Waving Action Recognition to FPGAs for Service Robots

We propose a system that offloads hand-waving action recognition processing in service robots to a field-programmable gate array (FPGA) accelerator. The accelerator implements a LUTNet-based reservoir computing (LUTNet-RC) model. Conventionally, such processing was performed by laptops connected to robots. Nevertheless, the proposed system offloads some of these task processing to an FPGA. This improves the overall efficiency of the robot and optimizes the use of computational resources. The experimental results show that in hand-waving action recognition, the LUTNet-RC on FPGA achieves higher accuracy and precision than the conventional echo state network (ESN) model, while significantly reducing the power consumption. In addition, the system can be expanded to large-scale networks, which are challenging to implement on laptops.