Hierarchical Graph-Based Planning with Interactive Object Awareness for Embodied Agents

To enable robust navigation among interactive objects, this study presents a hierarchical graph-based planning framework that incorporates object awareness into both global route selection and local motion planning. The global planner operates on a region connectivity graph, while the local planner leverages an interaction-aware object graph encoding pushable and passable objects. The two levels are coordinated through a cross-scale message-passing mechanism. Experiments are conducted on Gibson and AI2-THOR benchmarks with over 6,000 task episodes involving object manipulation during navigation. The proposed method achieves a 19.1% improvement in task completion rate and shortens planning time by 13.5% relative to non-hierarchical graph planners, highlighting the benefit of explicit hierarchical reasoning in interactive navigation scenarios.