EgoNav: Bridging Learned Waypoints and Geometry-Aware Local Control for Robust Indoor Navigation

Jing Wang; Shiqi Zhao; Hairong Qu; Peng Yin

cs.ROAug 26, 2026

EgoNav: Bridging Learned Waypoints and Geometry-Aware Local Control for Robust Indoor Navigation

Authors: Jing Wang, Shiqi Zhao, Hairong Qu, Peng Yin

Organizations: City University of Hong Kong, Hong Kong, China

Abstract

Image-goal navigation using lightweight topological maps is a practical paradigm for indoor robot deployment: the map requires only geotagged images, and localization relies on visual matching rather than precise pose estimation. However, learned waypoint predictors can produce targets that violate geometric constraints or deviate from the global path. Executing these waypoints safely further requires a local planner capable of collision avoidance, yet existing systems either lack one or rely on fixed parameters that cannot adapt to confined spaces. To address these limitations while retaining the navigational intuition of the learned predictor, we present EgoNav, a hierarchical system that implements this idea by generating candidates from semantically segmented traversable regions and scoring them alongside the learned waypoint for geometric safety, directional coherence, and fidelity to the learned prior. An adaptive local path planner then executes the refined waypoint with parameters modulated based on the refinement outcome. Experiments in Habitat-sim and on a physical humanoid robot show that EgoNav consistently outperforms contemporary baselines in both success rate and path efficiency.

Figures & tables

Fig. 1 : System workflow. An offline-collected image sequence serves as a topological map. A human instruction specifies the target, from which the shortest path to the goal is planned. An image-goal navigation model produces an initial waypoint, refined by our refinement module for higher-quality guidance.

Fig. 2 : The EgoNav algorithm pipeline. At timestep $t$ , the system receives the RGB image $I_{c}^{t}$ and depth image $D_{c}^{t}$ as inputs. The VPR module encodes $I_{c}^{t}$ and matches it against a pre-collected image sequence to retrieve a subgoal. This subgoal, together with the current image, is fed into GNM to produce an initial waypoint $\hat{w}_{t}$ . The candidate set $\mathcal{W}_{t}$ , which includes $\hat{w}_{t}$ , is then evaluated and refined by the waypoint refinement module through multi-criteria scoring. The corrected waypoint $w_{t}^{*}$ is passed to an adaptive path planner to generate the optimal trajectory $\tau_{t}^{*}$ .

Fig. 3 : Physical platform and interface. Left: the humanoid robot used for real-world validation. Right: the current query image $I_{c}^{t}$ , VPR-retrieved subgoal $I_{s}^{t}$ , and GNM-predicted waypoint (blue); the depth map (green); and collision-free paths from our local planner (yellow).

Parameter	Value	Derivation
$d_{\text{safe}}$	0.8 m	Robot footprint + margin
$d_{\text{prior}}$	1.6 m	$2\,d_{\text{safe}}$
$j_{\max}$	4	Max plausible advance per cycle
$\lambda_{s}$	0.5	Grid search
$\lambda_{c}$	1.5	Grid search
$\lambda_{p}$	1.0	Grid search

TABLE I : Parameter settings of EgoNav.

Fig. 4 : Waypoint refinement in action. Each row shows a navigation trajectory. The first column shows the subgoal image; the second shows semantic segmentation with skeleton extraction and candidate sampling. The following four columns show consecutive timesteps. Orange dots indicate the initial waypoint was selected; blue dots indicate a skeleton-sampled candidate was selected instead. The transition from orange to blue across frames reflects the robot turning to follow the refined direction.

Method	$<$ 10 m		10–20 m		$>$ 20 m
Method	SR	SPL	SR	SPL	SR	SPL
GNM [ 7 ]	63.3	61.0	36.7	34.5	26.7	24.2
ViNT [ 8 ]	60.0	57.8	33.3	31.0	20.0	17.6
PlaceNav [ 12 ]	70.0	67.4	46.7	44.0	36.7	33.1
NoMaD [ 31 ]	63.3	60.5	33.3	30.2	13.3	11.5
DeepExplore [ 11 ]	23.3	21.6	10.0	9.2	0.0	0.0

TABLE II : Comparison with baselines in Habitat-sim (MP3D), 30 episodes per method per difficulty level.

Method	$<$ 10 m		10–20 m		$>$ 20 m
Method	SR	SPL	SR	SPL	SR	SPL
GNM [ 7 ]	50.0	47.6	26.7	24.5	16.7	14.3
ViNT [ 8 ]	46.7	44.0	23.3	21.2	16.7	11.0
PlaceNav [ 12 ]	50.0	47.5	36.7	32.8	26.7	23.4
NoMaD [ 31 ]	50.0	45.5	26.7	24.1	16.7	13.8
DeepExplore [ 11 ]	16.7	15.3	6.7	6.0	0.0	0.0

TABLE III : Comparison with baselines in real-world environments, 30 episodes per method per difficulty level.

Fig. 5 : Qualitative comparison of navigation paths in simulation. The task requires navigating out of a narrow room through a doorway. EgoNav finds a collision-free path, while GNM and DeepExplore collide with obstacles (orange starburst) shortly after starting. VLFM avoids collisions but becomes trapped in a corner.

Configuration	SR ( $\uparrow$ )	SPL ( $\uparrow$ )
GNM	36.7	34.5
+ VPR retrieval (PlaceNav)	46.7	44.0
+ Local path planner (PlaceNav + Falco)	56.7	50.2
+ Waypoint refinement	70.0	61.2
+ Adaptive Planner Modulation	70.0	62.6
+ Localization filtering & relocalization (Full)	76.7	64.9

TABLE IV : Progressive ablation of EgoNav components on the 10–20 m navigation task (30 runs).

Configuration	SR ( $\uparrow$ )	SPL ( $\uparrow$ )
Full refinement	76.7	62.4
w/o $E_{\text{coher}}$	56.7	48.0
w/o $E_{\text{prior}}$	63.3	54.5
w/o $E_{\text{safe}}$	76.7	59.8
w/o refinement entirely	53.3	47.9

TABLE V : Ablation of individual refinement criteria on the 10–20 m navigation task (30 runs).

Component	Preprocess	Waypoint Prediction	Refinement	Planner
Time (ms)	100	20	40	2.5

TABLE VI : Average latency of core components on the humanoid platform.

Fig. 6 : Dynamic-obstacle avoidance (top: real; bottom: sim). In simulation, faded copies mark obstacles’ past positions.

		PlaceNav+Falco		EgoNav (Ours)
	Setting	SR	SPL	SR	SPL
Sim	Static	56.7	50.2	76.7	64.9
Sim	Dynamic	40.0	32.8	63.3	50.4
Real	Static	50.0	43.1	70.0	57.3
Real	Dynamic	30.0	20.8	50.0	36.5

TABLE VII : Effect of dynamic obstacles (10–20 m). Sim static rows are taken from the 10–20 m results in Table II. Sim: 30 episodes; real: 10 newly collected episodes over 5 scenes.

Robot Navigation

Robot navigation studies how robots perceive, localize, plan, and move safely toward goals in static or changing environments.

See all 536 papers

SemGeoNav:A Safety-Guided Visual Navigation Approach with Semantic Reasoning and Geometric Planning
Jun 15, 2026Yu Liu, Zongyang Chen, Yan Guo +2Robot Navigation Visual Robot Navigation
College of Intelligence Science and Technology, National University of Defense Technology, Changsha 410073, China.
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.
AgniNav: Configuration-Driven Cross-Embodiment Local Planning for Robot Navigation
Jun 9, 2026Tianhao Zang, Siwei Cheng, Haidong Huang +2Robot Navigation Safe Robot Navigation
Eastern Institute of Technology, Ningbo, China · University of Nottingham, Nottingham, UK · College of Information Science and Technology, Eastern Institute of Technology, Ningbo, P. R. China. +3

Jun 15, 2026J/K move · Enter open · S save

SemGeoNav:A Safety-Guided Visual Navigation Approach with Semantic Reasoning and Geometric Planning

Yu Liu, Zongyang Chen, Yan Guo, Chao Liu+1

College of Intelligence Science and Technology, National University of Defense Technology, Changsha 410073, China.

Robot Navigation Visual Robot Navigation Safe Robot Navigation Robot Motion Planning Robot Collision Avoidance

Open paper AskPDF arXiv

Learning-based visual navigation has enhanced semantic goal-reaching capabilities. However, due to their black-box nature, purely end-to-end models often lack explicit geometric constraints, leading to unpredictable and unreliable obstacle avoidance in open environments. Conversely, traditional geometric planners ensure safety but struggle with high-dimensional visual targets. To address these limitations, we propose SemGeoNav, a novel hierarchical visual navigation framework.It tightly integrates the high-level semantic reasoning of end-to-end models with the reliable local planning ability of geometry-based methods, achieving robust image-based navigation while significantly improving obstacle avoidance. Furthermore, we introduce a temporal trajectory smoothing mechanism to ensure continuous and stable robot motion. We evaluated SemGeoNav on a Unitree Go2 quadruped robot in real-world environments. The results demonstrate that SemGeoNav outperforms existing representative methods, including ViNT and NoMaD, achieving higher success rates and shorter navigation times.

Abstract

Robot Navigation

Explore similar work

SemGeoNav:A Safety-Guided Visual Navigation Approach with Semantic Reasoning and Geometric Planning

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

AgniNav: Configuration-Driven Cross-Embodiment Local Planning for Robot Navigation