Designing Socially Grounded Data Pipelines for Training and Operating Socially Intelligent Robots: Challenges and Future Directions

Designing socially intelligent robots presents a frontier challenge in human–robot interaction and information systems research, where technical architectures must align with the embodied, context-rich demands of social life. Current Vision–Language–Action (VLA) models integrate multimodal inputs but fall short in supporting socially coherent behavior, limiting their value as training pipelines. Using a clinical problematization approach, this study examines experimental pipelines to identify structural limitations in temporal coherence, continuity, multimodal integration, and affective inference. These shortcomings reveal a deeper misalignment between prevailing training paradigms and the sociotechnical requirements of interactional integrity. In response, we highlight perception–language models (PLMs) as a more socially attuned substrate, capable of extracting contextual signals and enhancing situational grounding for behavior modeling. We conclude by outlining a research agenda that advances spatiotemporal reasoning, affective modeling, and embodied coherence, thereby contributing to IS discourse on the design of trustworthy, socially adaptive robotic systems.