Martina Hasenjäger and Taizo Yoshikawa, "Machine learning approaches in human walk modeling", IROS 2019 Cutting Edge Forum "Human Movement Understanding for Intelligent Robots and Systems", 2019.

The combination of an increasing life expectancy and a low, decreasing birth rate has lead to aging societies in many countries. The resulting problems of a decreasing workforce and an increasing demand in health care are particularly acute in Japan. To address these problems, we aim to develop advanced physical assist devices to improve the quality of life and to extend activities of daily living and activities in the work place. We regard a human-in-the-loop approach a key technology: Assist devices are of greatest value if they provide support with the right timing, strength and quality. However, human motion varies from person to person as it depends on the person's condition, age, gender and health. Hence we need to develop intelligent cooperative assist robots that are able to adapt to the individual motion patterns and needs of the user. Our approach here is to analyze and understand human motion capture data using machine learning methods and to utilize the results to advance physical assist robotics. In this presentation, we will focus on human walk data recorded by inertial measurement units. We will discuss and compare machine learning approaches to heel strike estimation that either explicitly or implicitly model the temporal aspect of the data. We find that this task is solved best by recurrent neural networks and advocate the use of the currently non-standard model of echo state networks since the task is not too complex, the results are accurate, training is fast, and the model is easy to implement. Moreover, echo state networks can easily be extended to on-line learning and time series forecasting.