FactCheck: Feasibility-aware Long-term Action Anticipation with Multi-agent Collaboration

The article presents FactCheck, a multi-agent collaboration framework designed for long-term action anticipation (LTA) that employs a closed-loop "Observe-Plan-Verify" mechanism to enhance action feasibility. FactCheck features three specialized roles: an Observer that constructs a dual-form structured memory, a Planner that generates future action drafts based on historical data, and a Verifier that evaluates and refines these actions against a constructed History Action Graph. Experimental results on the EPIC-Kitchens-55 and EGTEA Gaze+ benchmarks indicate that FactCheck outperforms existing state-of-the-art methods, marking a significant advancement in feasibility-aware action anticipation for embodied intelligence applications.