Embedded systems and artificial intelligence for enhanced humanoid robotics applications

Abstract

This paper presents a method for collecting precise hand gesture (HG) data using a low-cost embedded device for an embedded artificial intelligence (EAI)-based humanoid robotics (HR) application. Despite advancements in the field, challenges remain in deploying cost-effective methods for accurately capturing and recognizing body gesture data. The ultimate objective is to develop humanoid robots (HRS) capable of better understanding human activities and providing optimal daily life support. In this regard, our approach utilizes a Raspberry Pi Pico microcontroller with a 3-axis accelerometer and a 3-axis gyroscope motion sensor to capture real- time HG data, describing ten distinct real-world tasks performed by the hand in experimental scenarios. Collected data is stored on a personal computer (PC) via a micro-python program, forming a dataset where tasks are classified using ten supervised machine learning (SML) models. Two classification experiments were conducted: the first involved predicting raw data, and the second applied normalization and feature extraction (FE) techniques to improve prediction performance. The results showed promising accuracy in the first phase (89% max), with further improvements achieved in the second phase (94% max). Finally, by employing similar methods, we can integrate highly trained machine learning (ML) models into embedded humanoid robotic systems, enabling real-time human assistance.