A continuous journey across environments

I2O2I-Nav

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

Zhengjie Shu1,2,* Wanyu Li2,* Di Xiao2,* Lingxin Jiang5 Lai Jiang2 Kaidi Hu2 Xincheng Tang2
Wenkang Hu2 Haoran Wang2 Jinchuan Zhang1 Chang Chen1 Wei Li3 Jia Pan1,† Ruigang Yang2,4,†
1 University of Hong Kong 2 Shanghai Jiao Tong University 3 Nanjing University 4 Neowa 5 Independent Researcher

* Co-first authors (equal contribution).   † Corresponding authors.

Paper PDF Code Video BibTeX
Four-stage I2O2I mission: Exit, OutdoorNav, Entrance, and open-vocabulary object navigation, with an illustrative route.
Four stages, two indoor–outdoor transitions, one continuous mission. Click any figure to view it at full size.

Overview

A robot's journey does not end at the doorway.

Service and delivery robots need to leave buildings, traverse outdoor spaces, enter a destination building, and find an object inside. I2O2I-Nav evaluates how these capabilities compose into a single, continuous mission.

The benchmark pairs reconstructed real-world environments with aligned 3D Gaussian Splatting (3DGS) rendering and collision geometry. Its three-level evaluation progresses from independent navigation primitives to two-stage compositions and full indoor-to-outdoor-to-indoor missions. Each stage continues from its predecessor's actual endpoint.

20real-world composite scenes
115,602.9 m²total annotated area
1,374evaluation episodes
3levels of composition
A journey from one building to another, through stairs, outdoor routes, and a destination entrance. This trajectory-following demonstration illustrates the environment and task stages; it is not a MODULAR-Agent evaluation rollout. Playback-speed labels are retained from the source video.

From Real-World Scenes to Executable Missions

Real-world scans are processed into environments with visual reconstruction, collision geometry, navigation maps, and object- and area-level annotations. Building and floor labels define goals and boundary checks; evaluator annotations are withheld from navigation methods.

Benchmark construction pipeline from scanning to scene processing and annotation, alongside the three-level task hierarchy.
Benchmark construction and hierarchical task definition, reproduced from the paper.

Connected Indoor–Outdoor Environments

The 20 composite scenes span schools, offices, residences, shopping areas, and parks, connecting multiple buildings and floors. Explore sampled views and reconstructed mesh geometry below.

First collection of sampled views from reconstructed benchmark environments.
Scene views 01 · Sampled environment views.
Outdoor area annotations used to identify region types and pedestrian-permitted traversal.
Area and region annotations
Object-level annotations used to define language-specified navigation goals.
Object-goal annotations

Three Levels of Compositional Difficulty

Moving between stages means switching the active task while retaining the robot's actual position and orientation. Composed episodes run without simulator resets.

L1

One stage at a time

Independent primitives

200 episodes per primitive
01

Exit

Find an exit, physically cross the boundary, and stop in a valid outdoor region.

02

OutdoorNav

Reach an outdoor goal while following a reference route and respecting pedestrian-permitted regions.

03

Entrance

Locate an entrance, enter the target building, and stop at a valid indoor location.

04

OVON

Find a language-specified object and stop at a valid location near a matching instance.

L2

Carry the state forward

Two-stage compositions

200 episodes per composition

I2O Indoor to outdoor

Exit→OutdoorNav

O2I Outdoor to indoor

OutdoorNav→Entrance
L3

One continuous journey

The complete I2O2I mission

174 episodes full missions
01ExitLeave the building
02OutdoorNavTraverse outdoor space
03EntranceEnter the destination
04OVONFind the target object

All four stages must succeed.Two indoor–outdoor transitions · Three stage handoffs · No resets

MODULAR-Agent

Plan. Execute. Verify. Recover.

A vision-language model coordinates frozen navigation policies through a reusable skills library. State estimation informs workflow and subtask planning; execution feedback drives completion verification, stage handoffs, and bounded recovery.

MODULAR-Agent architecture: a VLM coordinates a skill library and navigation policy library through state estimation, workflow and subtask planning, and verification and replanning.
The architecture shows the supported module scope. In L2/L3 evaluation, OmniNav executes Exit, Entrance, and OVON; SocialNav executes OutdoorNav. Evaluated inputs and backends are specified in the paper.

Evaluation: Composition Remains Challenging

Using the same underlying policies, MODULAR-Agent raises I2O full success from 14.00% to 30.50% and O2I from 12.50% to 32.50%. Full four-stage success improves from 1.15% to 4.02%, while remaining low.

Evaluation uses a collision-constrained kinematic agent in Isaac Sim. All methods share task episodes, observations, action interfaces, success criteria, and execution budgets.

Level 1 · Independent primitives

L1 · Exit, Entrance, and OVON
Method Exit Entrance OVON
SR ↑ SPL ↑ TL (m) SR ↑ SPL ↑ TL (m) SR ↑ SPL ↑ TL (m)
Uni-NaVid 0.00 0.00 1.14 0.00 0.00 1.49 7.00 7.00 0.59
OmniNav 29.00 18.29 156.11 14.50 10.26 35.53 38.00 35.73 21.19
LightNav-0 20.50 12.56 11.22 12.50 10.91 15.43 18.50 15.60 15.08
MODULAR-Agent (O) 45.50 29.56 74.54 51.00 30.78 48.02 40.50 32.26 46.28
L1 · OutdoorNav
Method SR ↑ SPL ↑ DCR ↑
OmniNav-fast 70.00 66.15 54.50
SocialNav 74.00 70.23 79.39
MODULAR-Agent (O) 74.00 69.32 55.35
MODULAR-Agent (S) 76.00 70.60 80.06

200 episodes per primitive. SR, SPL, and DCR are percentages; TL is in meters. Bold values reproduce the paper’s best-result formatting.

Level 2 · Two-stage compositions

I2O · Exit → OutdoorNav
Method FSR ↑ CSR ↑ SPL ↑ RC ↑ DCR ↑ C₁ ↑ C₂ ↑
OmniNav + SocialNav 14.00 17.25 11.90 28.50 42.24 20.50 68.29
MODULAR-Agent (O + S) 30.50 36.50 23.41 51.32 57.96 42.50 71.76
O2I · OutdoorNav → Entrance
Method FSR ↑ CSR ↑ SPL ↑ RC ↑ DCR ↑ C₁ ↑ C₂ ↑
OmniNav + SocialNav 12.50 42.75 8.78 82.93 53.39 73.00 17.12
MODULAR-Agent (O + S) 32.50 52.00 22.27 82.43 51.92 71.50 45.45

200 episodes per composition. All values are percentages. C₁ and C₂ are stage-conditional success rates in task order.

Level 3 · Full mission

L3 · Exit → OutdoorNav → Entrance → OVON
Method Full mission Stage-conditional success
FSR ↑ CSR ↑ SPL ↑ RC ↑ DCR ↑ C₁ ↑ C₂ ↑ C₃ ↑ C₄ ↑
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

174 episodes. All values are percentages. C₁–C₄ follow Exit, OutdoorNav, Entrance, and OVON. Full success requires all four stages.

Metric definitions and policy assignments

SR: success rate. FSR: full success rate. CSR: mean fraction of stages completed as a valid ordered prefix. SPL: success-weighted path efficiency. TL: mean traveled distance in meters. RC: reference-route progress from the final position's planar projection. DCR: outdoor distance compliance rate. Ck: stage completion conditional on successful completion of all preceding stages.

L1 (O)/(S) denotes OmniNav/SocialNav executors; OutdoorNav (O) uses OmniNav-fast. L2/L3 (O + S) combines OmniNav for Exit, Entrance, and OVON with SocialNav for OutdoorNav. The fixed chain advances on executor completion proposals; MODULAR-Agent adds completion verification and bounded recovery. All table values except TL are percentages. Higher is better for these percentage metrics; TL is descriptive.

Where Do Missions Break Down?

Seeing outdoor space does not guarantee a valid exit; approaching a doorway does not guarantee entry. Visible objects can be missed, and repeated actions can leave a robot stuck. These cases motivate explicit completion verification and recovery.

Eight primitive failure cases: missed and premature stopping on Exit, missed visible objects on OVON, failed entry on Entrance, and stagnation in vegetation on OutdoorNav.
Representative L1 failures from the paper: Exit (a,b), OVON (c,d), Entrance (e–g), and OutdoorNav (h).

Stage-conditional evaluation reveals both early attrition and final-goal failures. For MODULAR-Agent, conditional success at the final OVON stage is 21.21%, leaving efficient and reliable object search a substantial challenge.

Citation

BibTeX
@misc{shu2026i2o2inav,
  title = {I2O2I-Nav: A Benchmark with a Modular Agent for
           Long-Horizon Indoor-to-Outdoor-to-Indoor Navigation},
  author = {Zhengjie Shu and Wanyu Li and Di Xiao and Lingxin Jiang
            and Lai Jiang and Kaidi Hu and Xincheng Tang
            and Wenkang Hu and Haoran Wang and Jinchuan Zhang
            and Chang Chen and Wei Li and Jia Pan and Ruigang Yang},
  year = {2026},
  note = {Preprint},
  url = {https://i2o2i-nav.github.io/}
}