Stubborn: A Strong Baseline for Indoor Object Navigation

A Survey of Object Goal Navigation

“…Their work won the 2020 Habitat ObjectNav Challenge. Luo et al [49] also use a semantic top-down grid map borrowed from [48]. In contrast to previous modular methods using trained policies, they just set one of the corners of the map as a long-term goal and get impressive results.…”

“…Modular methods consist of a mapping module representing the environment, a policy to generate a long-term goal, and a path planner to navigate to the goal. Because modular methods inherit from classical navigation, some works directly use the framework without learning [49], [59]. Later, zero-shot methods [65]- [72] were proposed to solve problems when new objects are encountered in real-world applications.…”

A Survey of Object Goal Navigation

“…Their work won the 2020 Habitat ObjectNav Challenge. Luo et al [49] also use a semantic top-down grid map borrowed from [48]. In contrast to previous modular methods using trained policies, they just set one of the corners of the map as a long-term goal and get impressive results.…”

“…Modular methods consist of a mapping module representing the environment, a policy to generate a long-term goal, and a path planner to navigate to the goal. Because modular methods inherit from classical navigation, some works directly use the framework without learning [49], [59]. Later, zero-shot methods [65]- [72] were proposed to solve problems when new objects are encountered in real-world applications.…”

AION: Aerial Indoor Object-Goal Navigation Using Dual-Policy Reinforcement Learning

“…When perception and localization are fully known and accurate, modular methods have proven to be a mature solution for 2D zero-shot ObjectNav [1], which typically combines high-level planning and low-level goal-reaching while maintaining a semantic map for exploration. To accelerate the search, subgoal selection can be improved by rule-based methods [12], foundation models [13], RL policies [14], or powerful LLMs [3], [4]. For pure visual servoing systems that lack LiDAR sensors, map-based methods are prone to unreliable indoor localization.…”

MVP-Nav: Multi-layer Value Map Planner Navigator

“…Early successful agents primarily relied on active depth sensing (e.g., RGB-D) to build explicit 2D/3D occupancy representations for path planning [7]. A classic representative is Stubborn [22], which effectively leverages geometric baselines to prioritize collision-free frontier exploration, demonstrating that structured physical occupancy is fundamental to reliable navigation. As the field progressed, the focus shifted from pure geometric obstacle avoidance to cognitive process modeling [4].…”

A Survey of Object Goal Navigation

“…Their work won the 2020 Habitat ObjectNav Challenge. Luo et al [49] also use a semantic top-down grid map borrowed from [48]. In contrast to previous modular methods using trained policies, they just set one of the corners of the map as a long-term goal and get impressive results.…”

“…Modular methods consist of a mapping module representing the environment, a policy to generate a long-term goal, and a path planner to navigate to the goal. Because modular methods inherit from classical navigation, some works directly use the framework without learning [49], [59]. Later, zero-shot methods [65]- [72] were proposed to solve problems when new objects are encountered in real-world applications.…”

AION: Aerial Indoor Object-Goal Navigation Using Dual-Policy Reinforcement Learning

“…When perception and localization are fully known and accurate, modular methods have proven to be a mature solution for 2D zero-shot ObjectNav [1], which typically combines high-level planning and low-level goal-reaching while maintaining a semantic map for exploration. To accelerate the search, subgoal selection can be improved by rule-based methods [12], foundation models [13], RL policies [14], or powerful LLMs [3], [4]. For pure visual servoing systems that lack LiDAR sensors, map-based methods are prone to unreliable indoor localization.…”

MVP-Nav: Multi-layer Value Map Planner Navigator

“…Early successful agents primarily relied on active depth sensing (e.g., RGB-D) to build explicit 2D/3D occupancy representations for path planning [7]. A classic representative is Stubborn [22], which effectively leverages geometric baselines to prioritize collision-free frontier exploration, demonstrating that structured physical occupancy is fundamental to reliable navigation. As the field progressed, the focus shifted from pure geometric obstacle avoidance to cognitive process modeling [4].…”

A Survey of Object Goal Navigation

“…Their work won the 2020 Habitat ObjectNav Challenge. Luo et al [49] also use a semantic top-down grid map borrowed from [48]. In contrast to previous modular methods using trained policies, they just set one of the corners of the map as a long-term goal and get impressive results.…”

“…Modular methods consist of a mapping module representing the environment, a policy to generate a long-term goal, and a path planner to navigate to the goal. Because modular methods inherit from classical navigation, some works directly use the framework without learning [49], [59]. Later, zero-shot methods [65]- [72] were proposed to solve problems when new objects are encountered in real-world applications.…”

AION: Aerial Indoor Object-Goal Navigation Using Dual-Policy Reinforcement Learning

“…When perception and localization are fully known and accurate, modular methods have proven to be a mature solution for 2D zero-shot ObjectNav [1], which typically combines high-level planning and low-level goal-reaching while maintaining a semantic map for exploration. To accelerate the search, subgoal selection can be improved by rule-based methods [12], foundation models [13], RL policies [14], or powerful LLMs [3], [4]. For pure visual servoing systems that lack LiDAR sensors, map-based methods are prone to unreliable indoor localization.…”

MVP-Nav: Multi-layer Value Map Planner Navigator

“…Early successful agents primarily relied on active depth sensing (e.g., RGB-D) to build explicit 2D/3D occupancy representations for path planning [7]. A classic representative is Stubborn [22], which effectively leverages geometric baselines to prioritize collision-free frontier exploration, demonstrating that structured physical occupancy is fundamental to reliable navigation. As the field progressed, the focus shifted from pure geometric obstacle avoidance to cognitive process modeling [4].…”