Proactive Assistance Agent with Online Goal Recognition

Abstract

Effective human–AI collaboration requires agents to do more than simply follow instructions; they must also infer and anticipate human intentions. This ability to understand and predict others’ goals is also important in decentralized multi-agent settings, where communication is limited. Although goal recognition has been widely studied in the planning community, its integration with downstream collaborative behaviors, such as determining when and how an agent should assist, remains underexplored. This paper presents a unified model-based framework that combines online goal recognition with a proactive assistance module, enabling an agent to provide efficient support without prior knowledge of another agent’s true objective. Experiments in a grid domain show our proposed approach significantly improves collaborative efficiency over existing algorithms and demonstrates strong robustness across diverse configurations. By integrating goal recognition with other decision-making components, this work establishes a solid algorithmic foundation and an extensible framework for the capabilities required in next-generation human–AI collaboration.