Fast Recognition and Control of Walking Mode for Humanoid Robot Based on Pressure Sensors and Nearest Neighbor Search

In this paper, we propose a nearest-neighbor multi-reference learning system for control of humanoid-robot movements, using real-time data from pressure sensors embedded in the robot feet, which is processed with parallelized pipeline architecture for high-speed recognition of actual surface conditions. A first nearest-neighbor (1-NN) classifier is used to recognize the most similar reference pattern in terms of the smallest Euclidean distance. Our proposed architecture achieves a classification time of about $2.4\boldsymbol{\mu} \mathbf{s}$ with a total power consumption of 8.53mW at 100 MHz operating frequency when implemented on a low-cost FPGA (Cyclone-V GX-Series). The analysis results are further useful for a next-generation-ASIC-based AI-chip design for a robust real-time robot-learning system.