Detection and Recognition: A Pairwise Interaction Framework for Mobile Service Robots

A research team from KTH Royal Institute of Technology has introduced a novel AI framework designed to give mobile service robots—like autonomous lawnmowers or cleaning bots—the ability to understand human social interactions. Published on arXiv, the paper argues that for robots operating in human-populated environments, fine-grained activity recognition is often unnecessary. Instead, the key perception problem is identifying coarse-grained interactions between people to support safe and socially aware navigation. The proposed solution shifts focus from complex, holistic group analysis to a more practical, pairwise approach, providing a minimal yet sufficient perceptual unit for robot decision-making.

The framework operates in two stages: first, it uses lightweight geometric and motion cues (like proximity and movement patterns) to detect candidate pairs of interacting humans. Second, a relation network classifies the type of interaction. Evaluated on the JRDB and Collective Activity datasets, the method achieved comparable accuracy to heavier appearance-based models but with significantly reduced computational overhead and a smaller model footprint. This efficiency makes it ideal for integration into real-time mobile robot navigation systems. The team also demonstrated zero-shot generalization on a dataset collected by a lawnmower, proving its practicality for real-world deployment. The code will be released soon, paving the way for more intuitive and safer service robots in public spaces.