MOON: Multi-Objective Optimization-Driven Object-Goal Navigation Using a Variable-Horizon Set-Orienteering Planner

TL;DR

MOON reframes object-goal navigation as a multi-objective, budget-aware planning problem that balances unexplored region discovery with exploitation of observed landmarks. It integrates a Query-Based Occupancy Map (QOM), a StructNav-inspired perception-and-scene pipeline, and a variable-horizon Set Orienteering Problem (SOP) to achieve globally coherent long-horizon navigation. A data-driven budget-return predictor and mode-switching mechanism steer the agent between exploration and exploitation, while a high-speed transformer-based neural planner distills the expert SOP solver to millisecond latency with near-expert quality. Evaluations in large-scale, procedurally generated indoor environments show substantial SPL gains over baselines, demonstrating the practicality of scalable, long-horizon, landmark-informed ON. The framework offers a general template for long-horizon decision-making in resource-constrained, semantically structured navigation and can extend to multi-robot and multi-target settings.

Abstract

This paper proposes MOON (Multi-Objective Optimization-driven Object-goal Navigation), a novel framework designed for efficient navigation in large-scale, complex indoor environments. While existing methods often rely on local heuristics, they frequently fail to address the strategic trade-offs between competing objectives in vast areas. To overcome this, we formulate the task as a multi-objective optimization problem (MOO) that balances frontier-based exploration with the exploitation of observed landmarks. Our prototype integrates three key pillars: (1) QOM [IROS05] for discriminative landmark encoding; (2) StructNav [RSS23] to enhance the navigation pipeline; and (3) a variable-horizon Set Orienteering Problem (SOP) formulation for globally coherent planning. To further support the framework's scalability, we provide a detailed theoretical foundation for the budget-constrained SOP formulation and the data-driven mode-switching strategy that enables long-horizon resource allocation. Additionally, we introduce a high-speed neural planner that distills the expert solver into a transformer-based model, reducing decision latency by a factor of nearly 10 while maintaining high planning quality.

Figures (7)

• Figure 1: ON setup. This study explicitly models two types of visual recognition modules: a long-range landmark detector and a short-range target detector. When the sensing range of the landmark detector substantially exceeds that of the target detector, the landmark visiting order optimization problem (LOOP) becomes a key factor in efficient navigation. Specifically, an observation path exemplified by plan B, which maximizes the information gain-to-cost ratio, is preferable to those represented by plan A or plan C.
• Figure 2: Two conflicting objectives. The crescent moon-shaped yellow region represents the area newly observed by the robot. Landmarks are depicted as stars, with the previously observed area detected by the landmark detector shown in white, and the unobserved area shown in black. While unvisited landmarks exist within the previously mapped region (white), potential unvisited landmarks may also be present in the unmapped region (black). Determining which region to prioritize constitutes an ill-posed and computationally challenging planning problem.
• Figure 3: Prototype system. Our prototype is based on StructNav [RSS23] rss2023, with modifications limited to the essential components of our proposed method. Figure 1(a) illustrates StructNav, whereas Figure 1(b) depicts our approach. The key differences between the two systems are highlighted in red.
• Figure 4: An oblique view from a high vantage point serves as prior information for the exploration task. The oblique view depicted at the top was employed in an alternative exploration task, namely floor cleaning [IROS01], originally presented in iros2001. Over 20 years ago, when the original study was conducted, this view image exceeded the capabilities of the image recognition systems available at that time. Consequently, the approach employed a procedure whereby the robot autonomously selected the widest-area view from its visual experience and transmitted it to a teleoperator. The teleoperator subsequently used this view to plan the exploration, with an example of the planned path indicated by the white line in the figure. The middle part illustrates how the image, combined with a textual prompt, was used as input to a recent large language model (LLM). The bottom part shows the response from the LLM. The results indicate that the LLM demonstrates a level of scene understanding sufficient to provide prior knowledge for robotic path planning. The primary remaining challenge pertains to exploration path planning, which is the focus of the present study.
• Figure 5: Variable-horizon incremental map (VHIM). This figure presents three additional examples. In each image, the robot's workspace of interest is indicated by a colored ellipse. The top image shows a prior map of a previously visited location. The middle image depicts an aerial view acquired by a teammate drone. The bottom image illustrates a wide-angle view captured prior to entering the workspace, specifically at an intersection. These images correspond to those presented in howard2003radish, CityNav2024, and IV2019, respectively.