StARS lets robots learn your preferences like Netflix recommends movies

Social appropriateness in human-robot interaction is highly subjective: what one person finds polite, another may see as intrusive. To address this, researchers from the University of Cambridge and collaborators proposed StARS, a framework that reformulates robot action selection as a recommender system problem. In this analogy, each annotator (user) is treated as a unique entity, scenes/contexts are items, and appropriateness scores over a set of candidate robot actions become the ratings to predict. StARS integrates collaborative filtering with learnable scene representations to generate user-specific appropriateness scores without requiring a complete redesign of the underlying robot model. It is model-agnostic, meaning it can be paired with various scene encoders and backbone architectures, making it easy to retrofit into existing systems.

The framework was evaluated on two socially aware robotics datasets: MannersDB+ and SocNav1. Results showed that StARS consistently outperformed baselines in predicting appropriateness scores, improving agreement with human annotators. The system also demonstrated robustness under sparse preference feedback—a realistic scenario where users provide only a few examples of their social norms. By personalizing actions at the individual level, StARS enables robots to behave in ways that align with each user's unique expectations, whether in a home, office, or public space. The code is open-sourced, and the work will be presented at IROS 2026.