F2A-ObjectNav: An Adaptive Agentic Architecture for Object-Goal Navigation

Seyed Ali Mirghasemi; Alireza Taheri

cs.ROOct 14, 2026

F2A-ObjectNav: An Adaptive Agentic Architecture for Object-Goal Navigation

Authors: Seyed Ali Mirghasemi, Alireza Taheri

Organizations: Social and Cognitive Robotics Laboratory, Mechanical Engineering Department, Sharif University of Technology, Tehran, Iran

Abstract

Object-goal navigation requires an embodied agent to explore an unfamiliar environment, identify a target object, navigate toward candidate instances, and reliably determine whether the observed object satisfies the mission goal. This work investigates this problem through a two-stage development process. In Stage I, a fixed modular framework is used to examine combinations of exploration, object detection, and target validation methods. The results reveal clear trade-offs among navigation performance, validation reliability, execution time, and model-related cost, with no single fixed configuration remaining preferable across all evaluated conditions. These observations motivate F2A-ObjectNav (Fixed-to-Adaptive Object-Goal Navigation), an adaptive agentic architecture in which a rule-based Supervisor coordinates specialized components for navigation and exploration, object detection, target validation, semantic guidance, and Mission Memory. The architecture uses event-driven recovery, strategy revision, candidate management, and conditional semantic guidance while retaining lightweight execution during normal operation. F2A-ObjectNav is evaluated over 25 episodes in each of AI2-THOR, RoboTHOR, and ProcTHOR, achieving success rate (SR) / success-weighted path length (SPL) values of 0.84/0.52, 0.40/0.22, and 0.44/0.37, respectively. In addition to task-level metrics, architecture-specific behavioral metrics are used to characterize semantic-guidance usage, strategy switching, recovery activation, memory-supported behavior, and decision stability. The results show different patterns of adaptive intervention across the three environments, with higher intervention frequencies observed in RoboTHOR and ProcTHOR than in AI2-THOR.

Figures & tables

Figure 1: Research progression from the fixed modular study to the proposed adaptive agentic architecture. (AI-generated schematic using ChatGPT.)

Figure 2: Architecture of the fixed modular framework used in Stage I.

Config.	Exploration	Detection	Validation
C1	Random	YOLO11	YOLO-World
C2	Frontier-based	YOLO-World	GPT-4o
C3	Frontier-based	BLIP-2	YOLO-World
C4	BLIP-2-guided	YOLO-World	GPT-4o
C5	GPT-4o-guided	YOLO-World	BLIP-2
C6	Manual control (human-reference condition)

Table 1: Composition of the fixed modular configurations evaluated in Stage I.

Figure 3: Stage I simulation environments and representative outputs: (a) simple single-room layout, (b) complex three-room layout, (c) occupancy map produced during frontier-based exploration, and (d) first-person observation showing a detected target candidate.

Metric	Symbol	Evaluation purpose
Success Rate	SR	Proportion of successfully completed missions
Success-weighted Path Length	SPL	Mission success weighted by path efficiency
Environmental Adaptability Score	EAS	Relative retention of SR and SPL from the simple to the complex layout
Task Time	TT	Average duration from mission initialization to termination
Validation Accuracy	VA	Correctness of candidate-validation decisions
API Cost	AC	Average model-related inference cost per episode

Table 2: Metrics used to evaluate the Stage I configurations.

	SR		SPL
Config.	Simple	Complex	Simple	Complex	EAS
C1	0.73	0.27	0.51	0.09	0.273
C2	0.87	0.67	0.60	0.25	0.593
C3	0.87	0.60	0.58	0.29	0.595
C4	1.00	0.87	0.76	0.54	0.790
C5	1.00	0.80	0.71	0.51	0.759

Table 3: Navigation effectiveness of the fixed configurations in the simple and complex layouts.

	TT (s)		API Cost ($)
Config.	Simple	Complex	Simple	Complex	VA
C1	211	304	0.0000	0.0000	0.73
C2	105	281	0.0015	0.0015	0.91
C3	93	311	0.0000	0.0000	0.73
C4	74	367	0.0015	0.0015	0.91
C5	79	383	0.0031	0.0079	0.84

Table 4: Efficiency, model-related cost, and validation results of the Stage I configurations.

Figure 4: Comparative results of Stage I. Top: SR and SPL for each configuration, with Task Time shown on a secondary axis. Bottom: EAS and Validation Accuracy (left axis), with API Cost averaged equally across the two layouts on the right axis. EAS is calculated from Table 3 .

Stage I limitation	Stage II design response
One configuration remains active throughout the mission	Rule-based supervisory coordination and conditional strategy revision
Different failure events return to the same predefined execution process	Event-specific recovery and mission-level decision logic
Rejected candidates and unsuccessful search history are not systematically reused	Mission Memory and memory-informed candidate management
Semantic reasoning is determined by the selected configuration rather than current need	Conditional activation of the Semantic Advisor

Table 5: Design transition from the limitations of the fixed framework to the adaptive mechanisms introduced in Stage II.

Figure 5: Overall architecture of F2A-ObjectNav. A rule-based Supervisor coordinates specialized ObjectNav components and conditionally activates memory, recovery, strategy revision, and semantic guidance according to mission-level events. (AI-generated schematic using ChatGPT.)

Component	Main input	Main output	Primary function
Supervisor Agent	Mission state and execution events	High-level coordination decision	Selects the appropriate mission response using explicit rules and event conditions
Navigation and Exploration Agent	Pose, map, current goal, and mission state	Navigation actions and exploration subgoals	Explores the environment, approaches candidates, and reports navigation progress
Object Detection Agent	Visual observations and target representation	Candidate detections	Identifies possible target instances without declaring mission success
Target Validation Agent	Candidate observation and complete target description	Validation result	Determines whether an approached candidate satisfies the intended goal
Semantic Advisor Agent	Target, recent observations, and mission context	High-level semantic guidance	Suggests promising regions or search directions when conventional search is insufficient
Mission Memory	Mission events and intermediate outcomes	Stored episode-level records	Preserves information relevant to later candidate, exploration, recovery, and supervisory decisions

Table 6: Functional organization of the main F2A-ObjectNav components.

Figure 6: Supervisor Agent and its principal monitoring, information, and adaptive coordination roles. (AI-generated schematic using ChatGPT.)

Figure 7: Navigation and Exploration Agent, including its main responsibilities and information flow. (AI-generated schematic using ChatGPT.)

Figure 8: Object Detection Agent and its role in generating candidate target observations for subsequent inspection and validation. (AI-generated schematic using ChatGPT.)

Figure 9: Target Validation Agent and its role in assessing approached candidates against the complete mission goal. (AI-generated schematic using ChatGPT.)

Figure 10: Semantic Advisor Agent, which provides conditional high-level contextual guidance without directly controlling low-level navigation. (AI-generated schematic using ChatGPT.)

Figure 11: Mission Memory and its role in preserving mission-relevant information for later supervisory decisions. (AI-generated schematic using ChatGPT.)

Adaptive mechanism	Typical activation condition	Role in the architecture
Strategy switching	Insufficient progress, repeated failure, or ineffective current exploration behavior	Revises the active exploration strategy using the alternatives available to the Supervisor
Semantic-guided exploration	Prolonged search, repeated exploration, unsuccessful recovery, or insufficient semantic evidence	Requests high-level contextual guidance about a promising region or direction without directly controlling robot motion
Candidate management	Candidate detection, rejection, loss, or re-observation	Maintains candidate status and uses Mission Memory to avoid repeated evaluation of rejected instances
Spatial candidate merging	Multiple detections correspond to approximately the same estimated location	Consolidates repeated observations into a single candidate record
Recovery	Navigation stagnation, repeated trajectory, or failure to reach the current navigation objective	Replans, changes viewpoint or subgoal, clears an ineffective candidate, or resumes exploration
Goal-aware validation	The agent reaches a candidate with sufficient visual evidence	Evaluates the candidate against the complete target description before mission termination

Table 7: Adaptive mechanisms used in F2A-ObjectNav.

Figure 12: Representative scenes from the three simulation environments used for evaluation: AI2-THOR, RoboTHOR, and ProcTHOR.

Figure 13: Examples of generated maps during simulation, including the agent trajectory, intermediate goals, explored regions, and related navigation information.

Figure 14: Example of the object-detection process during simulation, showing representative YOLO11x (left) and YOLO-World v8x (right) outputs.

Figure 15: Example of the target-object validation process; the illustrated target object is an apple.

Component	Model	Role	Activation
Goal interpretation	GPT-4o mini	Extracts the primary target representation while preserving the complete description for validation	Mission initialization
Primary object detection	YOLO11x	Detects common object categories during routine perception	During normal search when applicable
Open-vocabulary detection	YOLO-World v8x	Detects target classes using flexible object names and accepted aliases	When open-vocabulary detection is required
Target validation	GPT-4.1 mini	Evaluates an approached candidate against the complete target description	After candidate approach
Semantic guidance	GPT-4o	Provides high-level contextual guidance about promising search regions or directions	On request from the Supervisor

Table 8: Models, roles, and activation modes used in F2A-ObjectNav.

Group	Metric	Purpose
Standard ObjectNav	Success Rate (SR)	Measures successful mission completion.
Standard ObjectNav	Success-weighted Path Length (SPL)	Measures mission success together with path efficiency.
System-level	Task Time (TT)	Measures average episode duration from initialization to termination.
System-level	Validation Accuracy (VA)	Measures the correctness of candidate-validation decisions.
System-level	API Cost (AC)	Measures the average inference cost of externally accessed models per episode.
Architecture-specific	Semantic Guidance Rate (SGR)	Measures the frequency of Semantic Advisor activation relative to episode length.

Table 9: Metrics used to evaluate F2A-ObjectNav.

Environment	SR (%)	SPL	TT (s)	VA (%)	API Cost ($)
AI2-THOR	84.00	0.52	48.63	89.5	0.0125
RoboTHOR	40.00	0.22	101.04	86.7	0.0273
ProcTHOR	44.00	0.37	109.73	84.2	0.0294

Table 10: Task- and system-level performance of F2A-ObjectNav across the three evaluation environments.

Figure 16: Task- and system-level results shown as grouped bars and lines. Bars show SR,

100\times

SPL, and VA on the left axis. The red line shows Task Time on the inner right axis, and the orange line shows API Cost on the outer right axis. Values correspond to Table 10 ; API Cost is expressed in USD per episode.

Environment	SGR (%)	SSR (%)	Recovery (%)	Memory (%)	Stability (%)
AI2-THOR	0.06	0.61	13.3	0.80	98.90
RoboTHOR	0.36	4.01	26.8	15.00	92.80
ProcTHOR	0.65	5.20	30.6	7.30	90.50

Table 11: Architecture-specific behavioral metrics of F2A-ObjectNav across the three evaluation environments.

Figure 17: Comparison of the architecture-specific behavioral metrics across the three evaluation environments. The legend label “Memory Usage” denotes Memory-Assisted Loop Response, as defined in Appendix B .

Mechanism	Aggregate diagnostic evidence	Execution-trace evidence	Observed functional role
Mission Memory	Memory-Assisted Loop Response was 0.80%, 15.00%, and 7.30% in AI2-THOR, RoboTHOR, and ProcTHOR, respectively.	In the recorded laptop mission, an incorrect candidate was rejected, stored in Mission Memory, and excluded from subsequent consideration before a later candidate was successfully confirmed.	Preserves episode-level information so that previous candidate and navigation outcomes can influence later decisions.
Recovery	Recovery Activation increased from 13.3% in AI2-THOR to 26.8% in RoboTHOR and 30.6% in ProcTHOR.	In the difficult remote-control mission, recovery was activated repeatedly after repeated states and navigation difficulties.	Provides event-triggered responses to stagnation, failed progress, and ineffective navigation states.
Strategy switching	SSR increased from 0.61% in AI2-THOR to 4.01% in RoboTHOR and 5.20% in ProcTHOR.	The recorded remote-control mission included switching between frontier-based and random exploration when the current search behavior became ineffective.	Prevents the architecture from remaining indefinitely committed to one exploration strategy when progress deteriorates.
Semantic Advisor	SGR remained low at 0.06%, 0.36%, and 0.65% in AI2-THOR, RoboTHOR, and ProcTHOR, respectively.	Semantic guidance was not required in the direct microwave-success mission but was invoked repeatedly in the more difficult remote-control execution.	Acts as a conditional source of high-level contextual guidance rather than a continuously active controller.
Candidate management and validation	Validation Accuracy remained between 84.2% and 89.5% across the three evaluation environments.	In the laptop mission, an initial candidate was rejected and stored before exploration continued and a later candidate was successfully confirmed.	Separates candidate generation from final target acceptance and prevents detector output from being treated directly as mission completion.
Supervisory coordination	Decision Stability was 98.90%, 92.80%, and 90.50% in AI2-THOR, RoboTHOR, and ProcTHOR, respectively.	Straightforward episodes proceeded with little intervention, whereas difficult executions involved recovery, strategy revision, memory interaction, and semantic assistance.	Coordinates adaptive intervention while retaining the current execution strategy when additional action is unnecessary.

Table 12: Execution-trace analysis of the principal adaptive mechanisms using evaluation-level behavioral measurements and recorded execution traces.

Environment	Method	Year	SR (%)	SPL
	DAT [ 12 ]	2023	82.39	0.489
	IOM [ 36 ]	2023	82.99	0.474
	MT [ 13 ]	2023	83.14	0.502
	Memory-MambaNav [ 22 ]	2025	83.19	0.498
AI2-THOR	F2A-ObjectNav (ours)	2026	84.00	0.520
	ESC [ 5 ]	2023	38.10	0.222

Table 13: Indicative cross-study comparison with selected ObjectNav methods. Values correspond to the protocols reported in the original studies and are not directly controlled across methods.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Item	Setting
Evaluation type	Simulation-based ObjectNav evaluation in indoor environments
Evaluation environments	AI2-THOR, RoboTHOR, and ProcTHOR
Mission specification	Object category or natural-language target description
Goal representation	Primary object representation for detection together with the complete target description retained for validation
Supervisory controller	Rule-based, event-driven Supervisor; no learned policy or LLM is used as the primary mission-level controller
Agent observations	Visual observations together with available depth, pose, and simulator-specific state information required by the navigation and supervisory components

Appendix

Table A.1 : Summary of the principal experimental settings used in the Stage II evaluation.

Environment	Role in the evaluation	Agent type	Main characteristic
AI2-THOR	Evaluation in structured residential scenes	Virtual first-person agent	Interactive indoor scenes including kitchens, living rooms, bedrooms, and bathrooms
RoboTHOR	Evaluation under robot-oriented navigation conditions	LoCoBot-based simulated agent	Indoor navigation with motion and perception constraints designed around a mobile-robot platform
ProcTHOR	Evaluation under greater structural and layout diversity	Virtual agent in procedurally generated environments	Procedurally generated houses with varied layouts, room connectivity, and object arrangements

Appendix

Table A.2 : Roles and main characteristics of the three Stage II evaluation environments.

Setting	Specification
Evaluation size	25 episodes per environment; 75 in total
Scene selection	Random selection from the available scenes of each environment
Target specification	Object category or natural-language description; target present, location withheld
Initialization	Valid starting pose; mission state and Mission Memory reset per episode
Agent inputs	RGB observations, depth, and navigation-related pose/state information
Target interpretation	LLM-based category extraction and preparation of a detailed validation description

Appendix

Table A.3 : Summary of the Stage II episode protocol.

Symbol	Term	Description
$C_{i}^{\mathrm{sem}}$	Semantic Guidance Call Count	Number of Semantic Advisor activations in episode $i$ .
$S_{i}^{\mathrm{sw}}$	Strategy Switch Count	Number of navigation or exploration strategy changes in episode $i$ .
$R_{i}^{\mathrm{ev}}$	Recovery Event Count	Number of recovery actions executed in episode $i$ .
$N_{i}^{\mathrm{issue}}$	Navigation Issue Count	Number of detected navigation problems in episode $i$ , including stagnation, repeated trajectories, candidate-approach failure, unproductive rotation, or prolonged lack of progress.
$M_{i}^{\mathrm{esc}}$	Memory-Supported Escape Count	Number of corrective or escape actions supported by Mission Memory in episode $i$ .
$N_{i}^{\mathrm{loop}}$	Navigation Loop Count	Number of detected trajectory-repetition or navigation-loop events in episode $i$ .

Appendix

Table B.1 : Symbols used in the architecture-specific behavioral metrics.

Component	Role	Model	Activation
Target Interpreter	Converts the mission goal into a detector-facing object class and a complete validation query	GPT-4o mini	Mission initialization
Semantic Advisor	Suggests a promising direction or region for continuing the search	GPT-4o	Conditional; used after insufficient progress, repeated navigation, or unsuccessful recovery
Attribute-Aware Vision Validator	Validates a candidate against the complete target description and explicit attributes	GPT-4.1 mini	Attribute-constrained target descriptions
Alias-Based Vision Validator	Validates a candidate using the target name and accepted equivalent terms	GPT-4.1 mini	Simple target categories without additional descriptive constraints
YOLO-World Class List	Defines the controlled target classes used for open-vocabulary detection	YOLO-World v8x	Open-vocabulary detection
YOLO11x	Performs primary detection for supported object categories	YOLO11x	Routine object detection

Appendix

Table C.1 : Summary of the prompts and controlled textual inputs used in F2A-ObjectNav.

Output	Role in the analysis
run_meta.json	Stores episode-level metadata such as environment, target, final status, recorded step count, execution time, and summary information.
flow_compact.json	Provides a compact structured sequence of the principal execution events and state transitions.
flow_compact.txt	Provides a human-readable summary of the mission execution flow.
appendix_bundle.json	Contains execution information organized for qualitative analysis and appendix reporting.
appendix_handoffs.txt	Records selected coordination events and information exchanges among system components.
exploration_map.png	Visualizes explored regions together with the trajectory recorded during the episode.

Appendix

Table D.1 : Principal outputs recorded during F2A-ObjectNav execution.

Property	Value
Environment	AI2-THOR
Scene	FloorPlan1
Target	Microwave
Final status	Successful
Recorded steps	Approximately 43
Execution time	Approximately 42 s

Appendix

Table D.2 : Summary of the direct successful execution for the microwave target.

Figure D.1: Representative direct successful execution for the microwave target in AI2-THOR.

Property	Value
Environment	AI2-THOR
Scene	FloorPlan201
Target	Laptop
Final status	Successful
Recorded steps	Approximately 63
Execution time	Approximately 67 s

Appendix

Table D.3 : Summary of the successful execution involving rejected-candidate management.

Figure D.2: Successful laptop-target mission after rejection and storage of an earlier candidate.

Property	Value
Environment	AI2-THOR
Scene	FloorPlan205
Target	Remote control
Final status	Unsuccessful
Recorded steps	Approximately 1,000
Execution time	Approximately 253 s

Appendix

Table D.4 : Summary of the unsuccessful remote-control mission involving multiple adaptive mechanisms.

Figure D.3: Snapshot during an unsuccessful remote-control mission involving strategy revision, recovery, semantic guidance, and Mission Memory; the displayed step precedes episode termination.

Property	Value
Environment	ProcTHOR
Scene	ProcTHOR-train-00057
Target	Soccer ball
Final status	Unsuccessful
Recorded steps	Approximately 1,000
Execution time	Approximately 405 s

Appendix

Table D.5 : Summary of the unsuccessful soccer-ball mission in ProcTHOR.

Figure D.4: Snapshot during an unsuccessful ProcTHOR mission in which no valid soccer-ball candidate was reached for final validation; the displayed step precedes episode termination.

Object-Goal Navigation

Object-goal navigation studies how robots find and approach requested objects in unfamiliar environments using visual, spatial, and language cues.

See all 86 papers

EffiNav: Fusing Depth and Vision-Language for Efficient Object Goal Navigation
Jun 17, 2026Zecheng Yin, Benedict Jun MaVLMs for Robotics Object-Goal Navigation
Systems Hub of Intelligence Transportation HKUST(GZ) Guangzhou, Guangdong
From Reactive to Map-Based AI: Tuned Local LLMs for Semantic Zone Inference in Object-Goal Navigation
Mar 9, 2026Yudai Noda, Kanji TanakaRobot Navigation Object-Goal Navigation
Yudai Noda and Kanji Tanaka are with the Graduate School of Engineering, University of Fukui, Japan.
What Matters in RL-Based Methods for Object-Goal Navigation? An Empirical Study and A Unified Framework
Oct 2, 2025Hongze Wang, Boyang Sun, Jiaxu Xing +5Robotic RL Robotic Perception
Computer Vision and Geometry Group, ETH Zurich, Switzerland · Robotics and Perception Group, University of Zurich, Switzerland · Robotic Systems Lab, ETH Zurich, Switzerland +2

Jun 17, 2026J/K move · Enter open · S save

EffiNav: Fusing Depth and Vision-Language for Efficient Object Goal Navigation

Zecheng Yin, Benedict Jun Ma

Systems Hub of Intelligence Transportation HKUST(GZ) Guangzhou, Guangdong

VLMs for Robotics Object-Goal Navigation

Open paper AskPDF arXiv

To locate a target object while exploring the unknown environment is a fundamental capability for autonomous agents, with applications ranging from search-and-rescue to field robots. A simplified version of such task is Object Goal Navigation (ObjNav). In ObjNav, successful arrival at the target object provides a basic measure of performance; however, the efficiency of the navigation trajectory is equally important, as it indicates how intelligently the agent explores and how much time remains for subsequent tasks. In unknown environments, the key to efficient navigation lies in deciding where to explore next. While many prior works aim to address this core challenge and achieved promising performance in certain settings, recent training-based models and non-training frameworks still suffer from generalization and efficiency issues respectively, which in the worst cases can lead to excessive exploration of already-visited areas or redundant back-and-forth motion. We evaluate EffiNav on two widely used simulation benchmarks Habitat Matterport 3D (HM3D) and Open-Vocabulary Object goal Navigation (OVON), and further validate its effectiveness on physical robots in real-world settings. We conduct failure analysis on massive simulation episodes. With minimal modification, we also extend EffiNav to a memory-augmented ObjNav task on the GOAT-BENCH dataset, demonstrating its adaptability beyond standard ObjNav settings. Across two standard metrics--Success Rate (SR) and Success weighted by Path Length (SPL), EffiNav matches or outperforms recent baselines, reflecting its efficiency, robustness, and practical applicability. Recognizing the different emphases of the two datasets, the performances reveals this framework is more balanced and generalizable for efficient ObjNav.

Abstract

Object-Goal Navigation

Explore similar work

EffiNav: Fusing Depth and Vision-Language for Efficient Object Goal Navigation

From Reactive to Map-Based AI: Tuned Local LLMs for Semantic Zone Inference in Object-Goal Navigation

What Matters in RL-Based Methods for Object-Goal Navigation? An Empirical Study and A Unified Framework