LOG-Nav: Efficient Layout-Aware Object-Goal Navigation with Hierarchical Planning

TL;DR

LOG-Nav addresses object-goal navigation in multi-room indoor environments by combining a global topology-based planner with a dense local ego-centric planner, all guided by an LLM-powered agent. The dual-level representation integrates a neural implicit scene function and a topometric map, enabling layout-aware routing without task-specific training or rewards. Empirical results on MP3D show substantial improvements in SR and SPL, and real-world deployments demonstrate practical robustness to dynamic changes. The approach offers a scalable, training-free paradigm for layout-aware navigation with broad applicability to household robotics and embodied systems.

Abstract

We introduce LOG-Nav, an efficient layout-aware object-goal navigation approach designed for complex multi-room indoor environments. By planning hierarchically leveraging a global topologigal map with layout information and local imperative approach with detailed scene representation memory, LOG-Nav achieves both efficient and effective navigation. The process is managed by an LLM-powered agent, ensuring seamless effective planning and navigation, without the need for human interaction, complex rewards, or costly training. Our experimental results on the MP3D benchmark achieves 85\% object navigation success rate (SR) and 79\% success rate weighted by path length (SPL) (over 40\% point improvement in SR and 60\% improvement in SPL compared to exsisting methods). Furthermore, we validate the robustness of our approach through virtual agent and real-world robotic deployment, showcasing its capability in practical scenarios.

Figures (8)

• Figure 1: Our proposed LOG-Nav. An efficient object-goal navigation approach with LLM-powered agent that realizes hierarchical planning based on topology and dense scene memory. The entire process is conducted automatically without costly training or complex rewards.
• Figure 2: Overview of our proposed method. The LLM agent takes user instructions as input and manages the optional action choices according to the prompts and data flow. Optional actions include exploring and recording the scene, constructing scene memory representation, planning, and executing navigation. Obstacles, error recognition, and iterative attempts are available.
• Figure 3: Hierarchical planning example. Global planning is conducted on the topological graph and generates $V=\{v_1, v_2, \dots , v_n\}$. Local planning, realized in IL approach, generates $W_{i,i+1}=\{w_i^1, w_i^2, \dots , w_i^m\}$.
• Figure 4: Data flow example of LLM Agent. Instructed by the user input, the agent conducts prompts, applies optional actions, and gets feedback from interaction.
• Figure 5: Navigation results on MP3D dataset. The left column shows global planning results, where gray points are entrance vertices in the topological map, red points show the starting positions and yellow points show the destinations. The right columns show local planning results on the ego-centric views.