Partially Observable Markov Decision Processes for Robotic Assistance in a Sensor-Fused Smart Home Environment

In this work the Partially Observable Markov Decision Process (POMDP) method is formulated for the determination of robot actions whose role is to assist a Human Activity Recognition (HAR) system in a smart home environment. The HAR consists of various sensors (wearable, microphones, cameras, smart plugs, temperature and humidity sensors) whose outputs are optimally fused to enhance the classification accuracy. In cases where the result of the HAR is uncertain a robot assists in the classification by approaching the person, performing its own recognition and providing, from its sensors, additional optical and audio data to the HAR. The POMDP model takes into consideration the precision of the HAR and assigns higher probability to state transitions that are more likely to occur.