Low-power real-time intelligent SoCs for smart machines

In this paper, we introduce low-power and real-time intelligent SoCs aimed at smart machines. To implement intelligent functions under low-power consumption, machine learning methods are tightly integrated with the traditional algorithms. At first, an object recognition processor (ORP) accelerating scale-invariant feature transform (SIFT) is presented with a visual attention based on convolutional neural network (CNN). For user interface (UI), a speech and gesture recognition processor (SGRP) based on convolutional deep belief network (CDBN) is presented with a voice activity detection (VAD) and a hand segmentation. At last, an artificial intelligence processor (AIP) for autonomous navigation is presented using A* tree search for path planning and reinforcement learning (RL) for dynamic obstacle avoidance. As a result, a prototype robot system integrating the presented SoCs is implemented and successfully demonstrated in the indoor environment.