<p>Domestic service robots need to have the cognitive ability to operate and succeed in complex, dynamic, and object-rich home environments. Inspired by human cognitive mechanisms combined with deep-learning techniques and soft computation, we developed a learning social referencing framework with mental imagery for domestic service robots and implemented the full architecture on a mobile manipulator robot, Fetch. This work demonstrates a comprehensive framework enabling service robots to address ambiguities in object selection tasks, and to continually learn under the guidance of a human interaction partner. We carried out a full system validation study with human participants to investigate user experience and attitudes towards the system, as well as the system’s functional success. We experimentally evaluated the proposed cognitive framework in four object selection scenarios and found the framework effective at enabling the service robot to be adaptive and capable of handling various ambiguities in interactions. Furthermore, participants perceived the robot positively in multiple dimensions, such as perceived intelligence, knowledge, sensibility and interactivity after interacting with the robot.</p>