An MRF-Based Intention Recognition Framework for WMRA with Selected Objects as Contextual Clues

To mitigate the physical burden of disabled people, we propose an approach that a robot could perceive the implied action intentions of disabled people by their selected objects. This article presents a framework for recognizing and learning human intentions based on selected household objects and interaction history. First, the intention network is modeled based on Markov random field (MRF) to connect the selected objects and daily activities. Second, the q-learning algorithm is added to provide the intention network with the function of adapting to the user’s intention preference. Then, we build the wheelchair-mounted robotic arms (WMRA) with a green laser pointer as human-robot interaction (HRI). Finally, we demonstrate the feasibility of the intention recognition framework by evaluating a scene comprised of objects from 11 categories, along with 7 possible actions, 36 single-object intentions, and 24 multiple-object intentions. We achieve approximately 70% reduction in fewer sessions than the Recursive Bayesian Incremental Learning and achieve approximately 87.5% and 86.2% reduction in interactions overall on recognizing multiple-object intention, respectively.