Cognitive Agentic AI: Probabilistic Novelty Detection for Continual Adaptation in HRI

Adapting to novel tasks in human-robot interaction (HRI) is crucial for long-term autonomy, yet remains a major challenge for autonomous agents deployed in unpredictable open-world settings. This paper introduces CAPA-AI, a novel framework that integrates probabilistic novelty detection with continual post-deployment adaptation achieved via transfer learning to address this challenge. The framework’s novelty detection component employs conditional probability and the Jaccard Index to identify unfamiliar tasks by quantifying their deviation from the agent’s knowledge base of previously learned tasks. Upon detecting a novel task, the agent utilises transfer learning to repurpose prior knowledge and update its models without retraining from scratch. We detail the design of CAPA-AI, including an isolated learning phase for initial skill acquisition and the construction of a dynamic knowledge base. The complete system was deployed on a social robot in real-world HRI scenarios to evaluate its performance. Experimental results demonstrated that the agent accurately detects novel tasks and adapts to them, achieving adaptation and novelty detection accuracies of 80% and 89%, respectively. These findings underscore the efficacy of the proposed approach and highlight a significant step towards robust open-world deployment of AI agents in HRI, where continuous adaptation and the safe handling of unforeseen tasks are essential.