Modeling, learning, and simulating human activities of daily living with behavior trees

Autonomy is a key factor in the quality of life of a person. With the aging of the population, an increasing number of people suffers from a reduced level of autonomy. That compromises their capacity of performing their daily activities and causes safety issues. The new concept of ambient assisted living (AAL), and more specifically its application in smart homes for supporting elderly people, constitutes a great avenue of the solution. However, to be able to automatically assist a user carrying out is activities, researchers and engineers face three main challenges in the development of smart homes: (i) how to represent the activity models, (ii) how to automatically construct theses models based on historical data and (iii) how to be able to simulate the user behavior for tests and calibration purpose. Most of recent works addressing these challenges exploit simple models of activity with no semantic, or use logically complex ones or else use probabilistically rigid representations. In this paper, we propose a global approach to address the three challenges. We introduce a new way of modeling human activities in smart homes based on behavior trees which are used in the video game industry. We then present an algorithmic way to automatically learn these models with sensors logs. We use a simulator that we have developed to validate our approach.