This project develops potentially groundbreaking methods that make service robots socially intelligent and capable of establishing durable relationship with their users. This relies on developing the capabilities to sense and express, which faces grand challenges: the low-quality signals and the poor context-awareness. We first propose a new paradigm called reinforcement fusion, which combines sensor signals in an interactive way: e.g. when a robot detects a sound direction, it turns towards the direction to see better and moves towards it to hear better. Reinforcement fusion is analogous to reinforcement learning, a known term in machine learning. It will dramatically improve robot's sensibility to the context including social behaviours. Secondly we propose a concept of social behaviour entrainment to adapt behaviours. Entrainment is the phenomenon that dialogue partners tend to adapt their speaking style of each other. Our hypothesis is that through reinforcement fusion based tracking and social information extraction, machine learning based context and user modelling, and social behaviour entrainment, durable social interaction between human and robot is achievable. The reinforcement fusion paradigm is applicable to all sorts of systems with steerable sensors.