NavArena: Automated Construction of Goal-Oriented Navigation Benchmarks from 3D Gaussian Splatting Reconstructions

Junhui Wang; Wei Yang; Xinyao Li; Ningjing Fan; Yuehao Yin; Xuecheng Chen; Chao Gao

cs.ROSep 4, 2026

NavArena: Automated Construction of Goal-Oriented Navigation Benchmarks from 3D Gaussian Splatting Reconstructions

Authors: Junhui Wang, Wei Yang, Xinyao Li, Ningjing Fan, Yuehao Yin, Xuecheng Chen, Chao Gao

Abstract

Fixed 3D Gaussian Splatting (3DGS) reconstructions provide realistic novel views but lack the traversability constraints, valid goals, and closed-loop protocols required for navigation evaluation. We introduce NavArena, an automated framework that transforms fixed 3DGS reconstructions into benchmarks for goal-oriented visual navigation. NavArena integrates a frozen 3DGS model for egocentric RGB-D rendering, an occupancy costmap derived from Gaussian density and height statistics for reachability and collision queries, and semantic goal candidates lifted from multi-view open-vocabulary masks. These components support the automatic generation and unified closed-loop evaluation of goal-oriented navigation episodes. Across more than 2{,}000 scenes, NavArena generates 22.2 million expert trajectories. Spatial and semantic evaluations assess the derived navigation representations, while policy rollouts demonstrate the diagnostic value of the unified evaluation protocol. NavArena enables scalable and reproducible navigation evaluation on large-scale 3DGS reconstructions, and all benchmark-generation tools, evaluation protocols, and derived assets will be released publicly.

Figures & tables

Platform	Rendering model	Collision model	Scale	Automated onboarding	Automated goal extraction	Goal-nav expert traj.	Closed- loop evaluation
Habitat + HM3D [ 2 , 12 ]	Mesh	NavMesh	1K	$\times$	$\times$	✓	✓
Habitat-GS [ 9 ]	3DGS	NavMesh	–	$\triangle$	$\times$	✓	✓
GS-Playground [ 13 ]	3DGS	Mesh	–	$\triangle$	$\times$	$\times$	$\times$
DISCOVERSE [ 14 ]	3DGS	Mesh	–	$\triangle$	$\times$	$\times$	$\times$
GaussGym [ 15 ]	3DGS	Mesh	–	$\triangle$	$\times$	$\times$	$\times$
NavArena	3DGS	Costmap	2K+	✓	✓	✓	✓

TABLE I: Comparison of navigation-environment construction pipelines based on the capabilities reported by their respective sources. NavArena focuses on automated benchmark construction from fixed 3DGS assets.

Fig. 1: NavArena construction pipeline. A frozen 3DGS model provides egocentric rendering, while separately inferred spatial and semantic layers define where the agent can move and which goals are valid. PointNav, ImageNav, and ObjectNav episodes are accepted only after reachability, collision, and goal-validity checks, then evaluated through the same closed-loop protocol.

Fig. 2: Training-free ObjectNav semantic grounding. Multi-view open-vocabulary masks are eroded, backprojected, fused by voxel voting, transferred to Gaussian primitives, and clustered into instance centers.

Query	CPU grid ( $\mu$ s)	FCL mesh ( $\mu$ s)
State	0.986	7.802
Motion	39.65	425.02

TABLE II: Spatial-layer query performance at

0.05\,\mathrm{m}

grid resolution; lower latency indicates cheaper closed-loop collision and motion-validity checks.

Resolution (m)	State FP (%)	State FN (%)	Motion FP (%)	Storage (KiB)
0.02	3.522	0.000	2.630	382.4
0.05	6.034	0.000	4.710	61.6
0.10	13.054	0.002	8.590	15.5

TABLE III: Occupancy-grid resolution trade-off against the FCL reference. FP and FN denote false-positive and false-negative rates; low FN is preferred because missed collisions create invalid episodes.

Method	Valid Center (%) $\uparrow$	Time/Scene (min) $\downarrow$
SceneSplat (Feedforward)	65.95	0.97
FlashSplat (Optimization)	57.36	135.31
Ours	92.25	6.56

TABLE IV: Semantic instance-center validity for ObjectNav goal construction.

Metric	PointNav/ImageNav	ObjectNav
Episodes	16.6M	5.6M
Mean Path Length (m)	7.35	3.51
Avg Steps	567	392
Total Duration (h)	130K	30K

TABLE V: Scale of the generated NavArena navigation dataset.

Task	Model	Dist. SR $\uparrow$	SPL $\uparrow$	NE $\downarrow$	CR $\downarrow$	AvgSteps
PointNav	LoGoPlanner	0.679	0.597	1.601	0.470	78.2
ImageNav	NaviBridger	0.434	0.389	2.095	0.694	694.2
	GNM	0.403	0.362	2.082	0.710	712.0
	NoMaD	0.385	0.364	2.160	0.707	734.9
	ViNT	0.288	0.286	2.203	0.728	806.0
ObjectNav	Uni-NaVid	0.374	0.175	2.491	0.183	163.6

TABLE VI: Aggregate distance-based zero-shot closed-loop performance on NavArena.

Fig. 3: Success degradation across navigation difficulty; longer optimal paths expose horizon-dependent failure modes under the same task interfaces.

Track	Model	Dist. SR $\uparrow$	Pose SR $\uparrow$	Pose Ret. $\uparrow$	Stop SR $\uparrow$
ImageNav	NaviBridger	0.434	0.255	0.588	$\mathrm{n/a}$
ImageNav	GNM	0.403	0.250	0.620	$\mathrm{n/a}$
ImageNav	NoMaD	0.385	0.189	0.491	$\mathrm{n/a}$
ImageNav	ViNT	0.288	0.180	0.625	$\mathrm{n/a}$
ObjectNav	Uni-NaVid	0.374	$\mathrm{n/a}$	$\mathrm{n/a}$	0.036
ObjectNav	PoliFormer	0.310	$\mathrm{n/a}$	$\mathrm{n/a}$	0.033

TABLE VII: Termination sensitivity under pose and stop criteria. Distance SR succeeds upon entering the target region without requiring

\mathrm{STOP}

; Stop SR requires

\mathrm{STOP}

inside the region.

Explore similar work

ReaDy-Go: Real-to-Sim Dynamic 3D Gaussian Splatting Simulation for Environment-Specific Visual Navigation with Moving Obstacles
Feb 12, 2026Seungyeon Yoo, Youngseok Jang, Dabin Kim +3Robotics Simulation Robot Navigation
Department of Aerospace Engineering, Seoul National University, Seoul, Republic of Korea · InnoCORE AI-Transformed Aerospace Research Center, KAIST, Daejeon, Republic of Korea
POINav: Benchmarking and Enhancing Final-Meters Arrival in Real-World Vision-Language Navigation
May 27, 2026Ruiyan Gong, Meisheng Zhang, Yuxiang Zhao +12Vision-Language Navigation
Amap CV Lab, Alibaba Group · Peking University
3D Gaussian Map with Open-Set Semantic Grouping for Vision-Language Navigation
May 26, 2026Jianzhe Gao, Rui Liu, Wenguan WangVision-Language Navigation 3D Scene Representation
The State Key Lab of Brain-Machine Intelligence, Zhejiang University

Feb 12, 2026, cs.ROJ/K move · Enter open · S save

ReaDy-Go: Real-to-Sim Dynamic 3D Gaussian Splatting Simulation for Environment-Specific Visual Navigation with Moving Obstacles

Seungyeon Yoo, Youngseok Jang, Dabin Kim, Youngsoo Han+2

Department of Aerospace Engineering, Seoul National University, Seoul, Republic of Korea · InnoCORE AI-Transformed Aerospace Research Center, KAIST, Daejeon, Republic of Korea

Robotics Simulation Robot Navigation Synthetic Data Generation for Robotics Robot Policy Generalization Dynamic Obstacle Avoidance

Open paper AskPDF arXiv

Visual navigation models often struggle in real-world dynamic environments due to limited robustness to the sim-to-real gap and the difficulty of training policies tailored to target deployment environments (e.g., households, restaurants, and factories). Although real-to-sim navigation simulation using 3D Gaussian Splatting (GS) can mitigate these challenges, prior GS-based works have considered only static scenes or non-photorealistic human obstacles built from simulator assets, despite the importance of safe navigation in dynamic environments. To address these issues, we propose ReaDy-Go, a novel real-to-sim simulation pipeline that synthesizes photorealistic dynamic scenarios in target environments by augmenting a reconstructed static GS scene with dynamic human GS obstacles, and trains navigation policies using the generated datasets. The pipeline provides three key contributions: (1) a dynamic GS simulator that integrates static scene GS with a human animation module, enabling the insertion of animatable human GS avatars and the synthesis of plausible human motions from 2D trajectories, (2) a navigation dataset generation framework that leverages the simulator along with a robot expert planner designed for dynamic GS representations and a human planner, and (3) robust navigation policies to both the sim-to-real gap and moving obstacles. The proposed simulator generates thousands of photorealistic navigation scenarios with animatable human GS avatars from arbitrary viewpoints. ReaDy-Go outperforms baselines across target environments in both simulation and real-world experiments, demonstrating improved navigation performance even after sim-to-real transfer and in the presence of moving obstacles. Moreover, zero-shot sim-to-real deployment in an unseen environment indicates its generalization potential. Project page: https://syeon-yoo.github.io/ready-go-site/.

Abstract

Explore similar work

ReaDy-Go: Real-to-Sim Dynamic 3D Gaussian Splatting Simulation for Environment-Specific Visual Navigation with Moving Obstacles

POINav: Benchmarking and Enhancing Final-Meters Arrival in Real-World Vision-Language Navigation

3D Gaussian Map with Open-Set Semantic Grouping for Vision-Language Navigation