AgenticNav: A Hierarchical Multi-Agentic System for LLM-Driven Autonomous Problem-Solving in Robotics

Kurzfassung

A primary challenge in autonomous robotics is bridging the semantic-execution gap in environments where pre-defined plans fail. This paper presents AgenticNav, a cognitive architecture that achieves autonomous problem-solving through a hierarchical Multi-Agentic System (MAS) orchestrated by a Large Language Model (LLM). Within this framework, specialized agents handle environmental perception and task execution, while a stateful agentic hierarchy facilitates failure detection and strategic recovery. When an environmental impasse is encountered, it escalates the execution failure to a strategic level where the LLM re-evaluates the environment’s topological structure to autonomously generate a viable alternative route. Validated in both a high-fidelity simulation and on a physical robot, the system demonstrates the ability to move beyond rigid instruction-following toward true, goal-oriented autonomy. Our results prove that the MAS framework effectively facilitates real-time, reason-based replanning, allowing robots to autonomously resolve complex execution failures in unpredictable real-world settings.