I2O2I-Nav: A Benchmark with a Modular Agent for Long-Horizon Indoor-to-Outdoor-to-Indoor Navigation

Zhengjie Shu; Wanyu Li; Di Xiao; Lingxin Jiang; Lai Jiang; Kaidi Hu; Xincheng Tang; Wenkang Hu

cs.ROOct 13, 2026

I2O2I-Nav: A Benchmark with a Modular Agent for Long-Horizon Indoor-to-Outdoor-to-Indoor Navigation

Authors: Zhengjie Shu, Wanyu Li, Di Xiao, Lingxin Jiang, Lai Jiang, Kaidi Hu, Xincheng Tang, Wenkang Hu,

Organizations: University of Hong Kong · Shanghai Jiao Tong University · Independent Researcher · NeoWa Robotics ·

Abstract

Service and delivery robots often need to complete long-horizon navigation missions spanning indoor and outdoor environments. Existing embodied navigation benchmarks primarily evaluate isolated stages or partial cross-domain journeys, providing limited support for assessing how heterogeneous navigation capabilities compose under continuous execution. We introduce I2O2I-Nav, a physics-enabled benchmark of 20 real-world composite scenes spanning over 115,000 square meters across multiple buildings and floors, with aligned 3DGS rendering and collision geometry. Its three-level evaluation progresses from independent primitives to two-stage compositions and complete four-stage I2O2I missions. Each stage starts from its predecessor's actual endpoint, while stage-conditional metrics track completion and localize failures along the mission. We further propose the MODULAR-Agent that coordinates frozen navigation policies through completion verification, stage handoffs, and bounded recovery. Evaluation across 1,374 episodes highlights the difficulty of continuous composition. The MODULAR-Agent achieves 40.5-76% success on individual primitives, but full I2O2I success remains low. Compared with a fixed policy chain using the same underlying policies, it more than doubles success on both Indoor-to-Outdoor (I2O) and Outdoor-to-Indoor (O2I), while improving full-mission success from 1.15% to 4.02%. Stage-wise analysis shows that the MODULAR-Agent enables more missions to reach the final object-navigation stage, where conditional success remains low, revealing challenges in both long-horizon stage execution and completion decisions.

Figures & tables

Figure 1: An I2O2I mission with four stages: Exit, OutdoorNav, Entrance, and OVON.

Benchmark	Domain	Cross-floor	3DGS	Simulator	Robot dynamics	Task dims
R2R	Indoor	$\surd$	$\times$	Matterport3D Simulator	$\times$	1
RxR	Indoor	$\surd$	$\times$	Matterport3D Simulator	$\times$	1
VLN-CE	Indoor	$\surd$	$\times$	Habitat	$\times$	1
VLN-PE	Indoor	$\times$	$\surd$	GRUtopia / Isaac Sim	$\surd$	1
HM3D-OVON	Indoor	$\times$	$\times$	Habitat	$\times$	1
BridgeNav	Outdoor	/	$\times$	Generated-video evaluation	$\surd$	2

Table 1: Comparison of embodied navigation benchmarks. Cross-floor denotes navigation across levels (e.g., via stairs) in benchmark episodes, not merely multi-floor scene assets.

\surd

: included;

\times

: excluded; /: unverified. “—” indicates not reported or not applicable.

Figure 2: Overview of I2O2I-Nav. (a) Benchmark construction pipeline. Large-scale real-world indoor–outdoor environments are scanned and processed into executable navigation scenes with collision geometry and navigation maps, together with object-level and area-level semantic annotations. (b) Hierarchical task definition with increasing compositional complexity. Level 1 evaluates four primitive tasks: Exit, Outdoor Navigation, Entrance, and Open-Vocabulary Object Navigation (OVON). Level 2 evaluates two-stage compositions, Indoor-to-Outdoor (I2O) and Outdoor-to-Indoor (O2I), where the second stage continues from the state reached by the first. Level 3 composes all four primitives into a continuous long-horizon I2O2I mission.

Figure 3: Overview of the MODULAR-Agent. A VLM agent uses instructions, observations, and services to coordinate a skill library and a navigation policy library. State estimation informs workflow planning over Exit, OutdoorNav, Entrance, and OVON. Subtask planning decomposes each workflow into executable steps, illustrated by finding stairs, descending, and exiting a building. Selected navigation policies execute these steps and return feedback for verification and replanning. The simulator environment interface closes the loop by executing actions and returning observations.

Method	SR $\uparrow$	SPL $\uparrow$	TL	SR $\uparrow$	SPL $\uparrow$	TL	SR $\uparrow$	SPL $\uparrow$	TL	Method	SR $\uparrow$	SPL $\uparrow$	DCR $\uparrow$
	Exit			Entrance			OVON				OutdoorNav
Uni-NaVid	0.00	0.00	1.14	0.00	0.00	1.49	7.00	7.00	0.59	OmniNav-fast	70.00	66.15	54.50
OmniNav	29.00	18.29	156.11	14.50	10.26	35.53	38.00	35.73	21.19	SocialNav	74.00	70.23	79.39
LightNav-0	20.50	12.56	11.22	12.50	10.91	15.43	18.50	15.60	15.08	MODULAR-Agent (O)	74.00	69.32	55.35
MODULAR-Agent (O)	45.50	29.56	74.54	51.00	30.78	48.02	40.50	32.26	46.28	MODULAR-Agent (S)	76.00	70.60	80.06

Table 2: L1 evaluation. SR, SPL, and DCR are percentages (higher is better); TL is in meters and is descriptive. Agent suffixes (O)/(S) denote OmniNav/SocialNav executors, with OmniNav-fast used for OutdoorNav (O). Bold values indicate the best result per task and metric.

	I2O: Exit $\to$ OutdoorNav							O2I: OutdoorNav $\to$ Entrance
Configuration	FSR $\uparrow$	CSR $\uparrow$	SPL $\uparrow$	RC $\uparrow$	DCR $\uparrow$	$C_{1}\uparrow$	$C_{2}\uparrow$	FSR $\uparrow$	CSR $\uparrow$	SPL $\uparrow$	RC $\uparrow$	DCR $\uparrow$	$C_{1}\uparrow$	$C_{2}\uparrow$
OmniNav + SocialNav	14.00	17.25	11.90	28.50	42.24	20.50	68.29	12.50	42.75	8.78	82.93	53.39	73.00	17.12
MODULAR-Agent (O + S)	30.50	36.50	23.41	51.32	57.96	42.50	71.76	32.50	52.00	22.27	82.43	51.92	71.50	45.45

Table 3: L2 composition metrics (%).

C_{1}=N_{1}/N

and

C_{2}=N_{2}/N_{1}

follow each task’s stage order. CSR credits valid prefixes, RC includes partial reference-route progress, and DCR is outdoor-only. Counts and undefined-rate conventions are specified in Appendix D . Bold values indicate the best result per task and metric.

	Full mission					Stage-conditional success
Configuration	FSR $\uparrow$	CSR $\uparrow$	SPL $\uparrow$	RC $\uparrow$	DCR $\uparrow$	$C_{1}\uparrow$	$C_{2}\uparrow$	$C_{3}\uparrow$	$C_{4}\uparrow$
OmniNav + SocialNav	1.15	14.22	0.90	35.62	43.81	29.89	67.31	28.57	20.00
MODULAR-Agent (O + S)	4.02	26.15	1.08	47.79	56.05	45.40	79.75	52.38	21.21

Table 4: L3 composition metrics (%).

C_{k}=N_{k}/N_{k-1}

, with

N_{0}=N

, follows Exit, OutdoorNav, Entrance, and OVON. FSR requires all four stages. RC uses reference-route progress; DCR excludes indoor distance. Bold values indicate the best result per task and metric.

Figure 4: Representative L1 failures. Exit: (a) timeout without STOP after reaching outdoors; (b) premature STOP before crossing outside. OVON: (c,d) failure to recognize a visible vending-machine goal. Entrance: (e,f) premature STOP outside a doorway; (g) timeout without entering a visible entrance. OutdoorNav: (h) repeated forward motion into vegetation, resulting in stagnation.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: 3D Gaussian renderings of the scanned scenes.

Figure 6: Mesh views of the reconstructed geometry used for simulation and collision queries.

Scene	Area (m 2 )	Object goals	Environment type
Scene 1	17,027.6	8	School
Scene 2	5,350.7	8	School
Scene 3	4,565.1	13	School
Scene 4	4,508.1	10	School
Scene 5	4,566.8	6	School
Scene 6	5,080.6	15	School

Appendix

Table 5: Inventory of the 20 scanned scenes, with a total area of

115{,}602.9\,\mathrm{m}^{2}

Figure 7: Roadblock assets used to close unscanned or poorly reconstructed branches.

Figure 8: Navigation map after scene editing and roadblock placement, with an enlarged local view.

Figure 9: Top-down region annotation. Region type and pedestrian-access status are separate attributes.

Panel of Figure 10 — Figure 10: Examples of semantic goal annotation. Left: an object target (locker bank). Right: an area target (storefront entrance). Orange outlines mark the annotated targets.

Figure 11: Word cloud of annotated OVON goal names. Word size is not interpreted as a measured category or episode frequency.

Level	L1				L2		L3	All
Task	Exit	OutdoorNav	Entrance	OVON	I2O	O2I	I2O2I	Total
Episodes	200	200	200	200	200	200	174	1,374

Appendix

Table 6: Evaluation episode inventory: 200 episodes for each L1 and L2 task and 174 for L3, totaling 1,374. All episodes are used for evaluation.

Panel of Figure 12 — Figure 12: Robot embodiment across the four I2O2I navigation stages.

Policy	Registry entry and aliases	Stages	Role
OmniNav	omninav (omninav_3view)	exit, entrance, ovon	default executor for Exit, Entrance, OVON
OmniNav-fast	omninav_fast (omninav-fast)	exit, entrance, outdoornav	outdoor executor in the (O) configuration
SocialNav	socialnav (social-nav)	outdoornav	outdoor executor in the (S) configuration
Uni-NaVid	uninavid (uni-navid)	exit, entrance, ovon	baseline policy
LightNav-0	lightnav (lightnav0)	exit, entrance, ovon	baseline policy

Appendix

Table 7: Policy registry entries used by the evaluation. Registry status gates an entry at build time.

Item	Setting
Model	kimi-k2.6
Endpoint	/chat/completions under the configured base URL https://api.kimi.com/coding/v1
Decoding	thinking disabled; max_tokens $256$ for location judgment, $1024$ for planning and replanning, $512$ for verification; request timeout $60$ s
Images	Up to four RGB frames in fixed order: the step-start view, then the forward, left, and right views captured after the stop; JPEG with long edge $640$
Messages	One system prompt and one user message per call, with prior events entering as rendered text
Parser fields	step_completed, step_confidence, step_reason, parent_completed, parent_confidence, parent_reason, evidence, retry_guidance

Appendix

Table 8: Verifier configuration for the reported runs.

Budget	Value	Scope
Executed actions	$800$	episode
Steps per attempt	$200$	attempt
Wall clock	$900$ s	episode
Model requests	$50$	episode
Verification cadence	every $50$ steps	stage
Periodic verifications	$20$	episode

Appendix

Table 9: Execution and recovery budgets of the reference agent. Model requests count actual HTTP calls.

Indoor Robot Navigation

See all 28 papers

NavVerse: Benchmarking Indoor-to-Outdoor Embodied Navigation in Continuous Robot Simulation
Jul 22, 2026Junzhe Wu, Yue Hu, Zeyu Han +4Object-Goal Navigation Robotics Simulation
University of Michigan, Ann Arbor · Tsinghua University
Beyond Isolation: A Unified Benchmark for General-Purpose Navigation
May 10, 2026Samson Sun, Tianyi Yang, Tengyue Wang +6Embodied Navigation Cross-Embodiment Robot Learning
1AutoLab, SAI, Shanghai Jiao Tong University · 2Research Lab, Anyverse Dynamics
osmAG-Nav: A Hierarchical Semantic Topometric Navigation Stack for Robust Lifelong Indoor Autonomy
Mar 30, 2026Yongqi Zhang, Jiajie Zhang, Chengqian Li +2Robot Navigation Indoor Robot Navigation
School of Computer Science and Technology, ShanghaiTech University, 393 Huaxia Road, Shanghai, 201210, Shanghai, China.

Jul 22, 2026J/K move · Enter open · S save

NavVerse: Benchmarking Indoor-to-Outdoor Embodied Navigation in Continuous Robot Simulation

Junzhe Wu, Yue Hu, Zeyu Han, Po-Hsun Chang+3

University of Michigan, Ann Arbor · Tsinghua University

Object-Goal Navigation Robotics Simulation Robot Navigation Indoor Robot Navigation Embodied Navigation Vision-Language Navigation+2

Open paper AskPDF arXiv

Robots deployed in delivery, campus, and emergency-response settings often need to navigate from buildings to streets within a single continuous episode. Existing benchmarks usually evaluate indoor and outdoor navigation separately, and many abstract away robot execution, leaving exit finding, boundary traversal, adaptation, and kinodynamic failures underexplored. We introduce NavVerse, a physics-enabled benchmark for indoor-to-outdoor embodied navigation. NavVerse contains 100 indoor scenes, 50 urban outdoor scenes, and 50 indoor-to-outdoor scenes, and 10,000 episodes spanning Object Navigation, Vision-and-Language Navigation, and Place Navigation tasks, where agents search for semantic points of interest such as restaurants or banks. Agents are evaluated through executable robot interfaces using task-success, path-efficiency, and safety metrics. Zero-shot experiments with RL, VLA, and modular baselines show that current agents remain far from solving cross-context navigation: end-to-end VLAs obtain the highest zero-shot success, while the modular method provides the strongest safety profile. PlaceNav further reveals a clear drop from outdoor to indoor-to-outdoor scenes, indicating that adaptation remains major bottleneck.

Abstract

Indoor Robot Navigation

Explore similar work

NavVerse: Benchmarking Indoor-to-Outdoor Embodied Navigation in Continuous Robot Simulation

Beyond Isolation: A Unified Benchmark for General-Purpose Navigation

osmAG-Nav: A Hierarchical Semantic Topometric Navigation Stack for Robust Lifelong Indoor Autonomy