ScheNav: Schematic-Guided Object Goal Navigation

Guichen Huang; Chenchen Jing; Che Sun; Yuwei Wu; Yunde Jia

cs.ROOct 25, 2026

ScheNav: Schematic-Guided Object Goal Navigation

Authors: Guichen Huang, Chenchen Jing, Che Sun, Yuwei Wu, Yunde Jia

Organizations: Beijing Institute of Technology · Shenzhen MSU-BIT University · Zhejiang University of Technology

Abstract

Object goal navigation requires an embodied agent to locate and reach target objects in complex environments. Humans often rely on low-cost yet informative schematic maps, such as rough sketches or directional annotations. We develop a workflow for schematic-guided embodied navigation with two components: a training-data generation pipeline and ScheNav, a closed-loop navigation framework. Our lightweight data-generation pipeline turns 2D floorplans into navigation training environments, yielding Furnished-RPLAN, a dataset of over 33k furnished layouts with paired navigation episodes and noise-controllable schematic maps. ScheNav turns rough schematic guidance into exploration and stopping decisions by using SchemaField, a learned spatial module that fuses the guidance with accumulated observations and predicts spatial value fields. The framework updates these fields as exploration proceeds, with trajectory-based training and a cross-frame consistency objective supporting coherent guidance across changing viewpoints. ScheNav learns to use schematic guidance entirely from simulated floorplans and transfers this ability to unseen embodied environments. Experiments show improved navigation success and path efficiency across schematic maps, goal ellipses, and pointed directions, together with tolerance to the tested schematic perturbations and omissions. Together, these results demonstrate that rough schematic guidance offers a practical way to improve embodied navigation.

Figures & tables

Figure 1: Overview of schematic-guided navigation. A human provides imperfect schematic guidance, such as a rough sketch or a pointed direction. ScheNav grounds the guidance in the agent’s observations and predicts a value field that rises toward the goal. Guided by the predicted field, the agent reaches the goal along an efficient route in an environment it has barely explored.

Figure 2: The ScheNav framework. Schematic guidance at any level, a schematic map, ellipses, a pointed direction, or none, is encoded by SchemaField, the only learned component our framework introduces: a schematic-fusion transformer fuses the guidance with the observed map from the map memory, and a distance decoder and a corridor decoder predict the two value fields from the fused embedding in a single forward pass. Two rule-based components read the fields. The Goal Verification module compares the distance-field value at goal instances detected in the map memory against a threshold $\tau_{\mathrm{stop}}$ , and the decision transformation scores frontier candidates with the corridor field and gates stopping, emitting a waypoint or a stop. Map memory, frontier extraction, and execution are supplied by the host navigation system, and each new observation closes the loop.

Figure 3: Furnished-RPLAN construction pipeline. Starting from an RPLAN vector floorplan, a vision-language pipeline proposes and extracts a machine-readable furniture layout. The resulting furnished environment is used to generate navigation episodes, paired schematic guidance, and the local observations consumed by ScheNav.

Method	SR $\uparrow$	SPL $\uparrow$
VLFM [ 42 ]	14.9	9.3
GOAT [ 8 ]	17.0	8.8
PSL [ 29 ]	26.0	10.2
UniGoal [ 41 ]	20.2	11.4
Context-Nav [ 14 ]	26.2	9.1
Ours (no schematic input)	21.9	5.8

Table 1: Comparison with the state of the art on UniGoal’s instance-level HM3D benchmark (1000 episodes; SR and SPL in percent). Baseline numbers follow Context-Nav [ 14 ] . Baselines receive a textual goal description; all “Ours” rows use the unified 34k model with guidance at the indicated level.

Method	SR $\uparrow$	SPL $\uparrow$	sSPL $\uparrow$	DTS $\downarrow$
UniGoal (baseline)	17.2	9.5	21.3	5.65
+ single-frame field (incomplete)	15.6	9.4	22.5	5.26
+ Goal Verification on single-frame field (incomplete)	12.3	8.5	23.3	6.16
+ seq. + consist. + corridor	23.8	15.5	28.7	4.30
+ Goal Verification (complete)	45.9	27.1	38.7	2.53
Ours (all, $\sim$ 16k)	54.9	35.0	48.6	1.75

Table 2: Component progression on the frozen Habitat set (122 episodes, 36 scenes; SR, SPL, and sSPL (soft-SPL) in percent, DTS in meters). Rows marked “+” progressively modify UniGoal [ 41 ] and are trained on 800 layouts. The first two rows are incomplete combinations included to show that adding components independently can hurt; the complete Goal Verification row follows trajectory-consistent field training. The last two rows are the unified model trained on

\sim

16k and

\sim

34k layouts.

Corridor ablation (flip rate $\downarrow$ ):
Metric	Before	After (full)
Direction-flip rate ( $>90^{\circ}$ ) $\downarrow$	0.456	0.101
World-frame corr. (post-turn) $\uparrow$	$-0.15$	+0.80
Frontier ranking corr. $\uparrow$	0.35–0.48	0.750
Top-decision margin $\uparrow$	$1\times$	3.4 $\times$
plain corridor ( $\tau_{c}{=}60$ )	0.367

Table 3: Cross-frame consistency before and after the fix (paired counterfactual replay, 97 decisions) and the corridor ablation (direction-flip rate), measured on the component-progression generation.

Figure 4: Closed-loop trajectories on one episode under decreasing guidance. The five panels show our four input conditions and unmodified UniGoal on the same episode, rendered on a textured top-down view of the scene. The dashed inset in each panel shows the schematic input the agent receives, and is blank for the no-guidance condition and the baseline; in the direction panel the floorplan behind the arrow is drawn only for illustration, and the model receives the direction vector alone. Orange curves trace the agent from start to stop, a green check marks a successful stop, and the red cross marks a failed stop. The path lengthens gracefully as guidance weakens, from the schematic map through the ellipse and direction conditions to no guidance at all, while the host stops incorrectly after prolonged, tangled exploration.

Robot Navigation

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

See all 554 papers

FloVerse: Floor Plan-Guided Multi-Modal Navigation
Jun 12, 2026Weiqi Huang, Shuangyi Dong, Jiaxin Li +3Embodied Navigation Robot Imitation Learning
School of Computer Science & Technology, Beijing Institute of Technology
FloorPlan-VLN: A New Paradigm for Floor Plan Guided Vision-Language Navigation
Mar 18, 2026Kehan Chen, Yan Huang, Dong An +5Vision-Language Navigation Embodied Navigation
New Laboratory of Pattern Recognition (NLPR), Institute of Automation, Chinese Academy of Sciences and School of Artificial Intelligence, University of Chinese Academy of Sciences, China · AMap, Alibaba Group · BAAI +2
SkillNav: Score-Level Skill Intervention for Zero-Shot Object Goal Navigation
Jul 17, 2026Ruijie Sang, Yiqun Duan, Pinhan Fu +3Vision-Language Navigation Object-Goal Navigation
D-Robotics · HAI Center, University of Technology Sydney · School of Computer Science, Wuhan University

Jun 12, 2026J/K move · Enter open · S save

FloVerse: Floor Plan-Guided Multi-Modal Navigation

Weiqi Huang, Shuangyi Dong, Jiaxin Li, Yifei Guo+2

School of Computer Science & Technology, Beijing Institute of Technology

Embodied Navigation Robot Imitation Learning

Open paper AskPDF arXiv

Floor plans encapsulate compact spatial priors, enabling agents to navigate unseen scenes more efficiently. While prior work has explored floor plan-guided navigation, it has focused mainly on PointNav and a limited set of environments. To bridge this gap, we introduce FloVerse, a new task for floor plan-guided embodied navigation that unifies PointNav, ObjectNav, and ImageNav. To support FloVerse, we assemble FloVerse-1.6K, a large-scale dataset of 1.6K scenes from HM3D and Gibson 4+, paired with corresponding floor plans, comprising 240K expert trajectories and 12M RGBD frames. We further propose ThreeDiff, a two-stage imitation learning policy comprising a planner, a diffusion-based multimodal goal-reasoning module trained via masked-modality modeling, and a refiner, a depth-based trajectory-refinement module for safe execution. Extensive experiments demonstrate that (1) floor-plan priors improve navigation performance across all goal modalities, and (2) ThreeDiff implicitly captures spatial information from floor plans. These results underscore the effectiveness of spatial priors and validate our proposed unified approach for floor plan-guided embodied navigation.

Abstract

Robot Navigation

Explore similar work

FloVerse: Floor Plan-Guided Multi-Modal Navigation

FloorPlan-VLN: A New Paradigm for Floor Plan Guided Vision-Language Navigation

SkillNav: Score-Level Skill Intervention for Zero-Shot Object Goal Navigation