SOR-Nav: Search or Relocate? Context-Gated Exploration and Cross-Region Relocation for Object Navigation

Yuan Ji; Zirui Li; Yuxin Cai; Shuge Wu; Boon Siew Han; Chen Lv

cs.ROSep 28, 2026

SOR-Nav: Search or Relocate? Context-Gated Exploration and Cross-Region Relocation for Object Navigation

Authors: Yuan Ji, Zirui Li, Yuxin Cai, Shuge Wu, Boon Siew Han, Chen Lv

Organizations: School of Mechanical and Aerospace Engineering, Nanyang Technological University, 639798, Singapore · Schaeffler Hub for Advance REsearch(SHARE) at NTU, Singapore

Abstract

Object navigation requires an embodied agent to find an object in an unseen environment under partial observability and a limited motion budget. Existing methods primarily optimize where the robot should go next by ranking candidate destinations. In contrast to these methods, we present SOR-Nav, a hierarchical navigation system that explicitly arbitrates between continuing to explore the current context and abandoning it for a more promising reachable region. First, an autonomous semantic exploration system is built that accumulates persistent 3D object clusters and organizes reachable frontiers into a cluster decision graph to provide an efficient search abstraction. Then, SOR-Nav uses a context-gated LLM-driven object-search supervisor to evaluate the suitability of the current search context and decide whether to continue exploration or perform cross-region relocation to another reachable frontier cluster. Across the complete, unfiltered validation sets of HM3D-v1, HM3D-v2, and MP3D, SOR-Nav achieves the strongest reported Success Rate (SR) and Success weighted by Path Length (SPL) on all three benchmarks. On MP3D in particular, it more than doubles the previous best SPL from 18.1% to 38.5% while increasing SR from 50.7% to 61.8%. Nested HM3D-v2 ablations validate the proposed decision structure, while a continuous three-target physical deployment demonstrates persistent ObjectNav operation in real-world scenarios.

Figures & tables

Fig. 1: Overview of the SOR-Nav object-search workflow. RGB and depth observations update persistent 3D object clusters and a cluster decision graph in the autonomous semantic exploration system. The LLM-driven supervisor prioritizes confirmed target evidence; otherwise, its context gate determines whether to continue exploration or relocate to a graph-reachable frontier cluster. All outcomes share the same mission executor, and subsequent observations and mission results close the loop.

Method	ZS	HM3D-v1		HM3D-v2		MP3D
Method	ZS	SR	SPL	SR	SPL	SR	SPL
L3MVN [ 2 ]	Yes	50.4	23.1	36.3	15.7	34.9	14.5
ESC [ 1 ]	Yes	39.2	22.3	–	–	28.7	11.2
VoroNav [ 6 ]	Yes	42.0	26.0	–	–	–	–
VLFM [ 3 ]	Yes	52.5	30.4	63.6	32.5	36.4	17.5
SG-Nav [ 5 ]	Yes	54.0	24.9	49.6	25.5	40.2	16.0

TABLE I: ObjectNav comparison based on the SysNav simulation table and SOR-Nav evaluation results. SR and SPL are reported in percentages.

Variant	$N$	SR $\uparrow$	SPL $\uparrow$	SL $\downarrow$	TO $\downarrow$
SOR-Nav (full)	1000	87.6	52.7	5.0	7.4
w/o relocation gate	1000	82.2	39.9	9.7	8.1
w/o relocation gate and TSP	1000	76.4	29.2	16.3	7.3

TABLE II: Nested ablation study on HM3D-v2. SR, SPL, step-limit (SL), and timeout (TO) rates are percentages;

N

is the number of evaluated episodes.

Fig. 4: Paired qualitative results for 18 successful episodes across HM3D-v2, HM3D-v1, and MP3D, with six target categories shown per dataset. Within each episode panel, the goal-facing terminal RGB observation is shown directly above its corresponding explored occupancy map and executed trajectory. The yellow overlay marks the detected target category; the title reports the target and true dataset episode index. Portrait RViz maps with width-to-height ratio below 0.95 are rotated $90^{\circ}$ clockwise for compact display.

Fig. 5: Representative events from a continuous real-world ObjectNav run in which the next request was issued after each successful target confirmation. Each labeled pair shows the first-person observation (top) and accumulated map, graph, and trajectory (bottom). The sequence highlights the relocation trigger, passage through the office doorway, and terminal observations for television, trash can, and signboard; all map panels use the final viewport. The reported distance is cumulative robot travel from the beginning of the sequence, rather than geodesic distance to a target.

Explore similar work

Rethinking Embodied Navigation via Relational Inductive Bias
Jun 9, 2026Weitao An, Chenghao Xu, Xu Yang +1Menu Open-Vocabulary
School of Electronic Engineering, Xidian University, Xi’an 710071, China · School of Information Science and Engineering, Hohai University, Nanjing 210098, China
IntentNav: Learning Spatial-Visual Object Navigation from Human Demonstrations
Jun 6, 2026Yuxin Cai, Zongtai Li, Maonan Wang +9Navigation Visual Memory
Nanyang Technological University · Carnegie Mellon University · The Chinese University of Hong Kong +1
EffiNav: Fusing Depth and Vision-Language for Efficient Object Goal Navigation
Jun 17, 2026Zecheng Yin, Benedict Jun MaObject Goal Navigation Vision-Language Navigation
Systems Hub of Intelligence Transportation HKUST(GZ) Guangzhou, Guangdong

Jun 9, 2026, cs.ROJ/K move · Enter open · S save

Rethinking Embodied Navigation via Relational Inductive Bias

Weitao An, Chenghao Xu, Xu Yang, Cheng Deng

School of Electronic Engineering, Xidian University, Xi’an 710071, China · School of Information Science and Engineering, Hohai University, Nanjing 210098, China

Menu Open-Vocabulary Inductive Bias

Open paper AskPDF arXiv

Object navigation requires an agent to locate a target in an unknown environment through visual observations. Existing methods typically rely on open-vocabulary detectors or vision-language models (VLMs) to answer where to search, but often overlook what not to trust - which semantic cues are unreliable. Open-vocabulary perception is prone to systematic misleading evidence: false positives, outdated static priors, and repeated failed exploration due to lack of embodied verification, which contaminates mapping and decision-making. Such errors are rooted in structured object relations in real-world scenes. To address this, we propose DB-Nav, a framework that reshapes the search space via dual relational biases. It factorizes target-centric relations into an Activation Bias (propagates contextual evidence) and an Inhibition Bias (suppresses unreliable regions via perceptual confusion and action-level falsification). These biases are unified into a Relational Activation-Inhibition Exploration Graph that modulates frontier exploration values using online observations and failed accesses. Experiments on ObjectNav benchmarks show that DB-Nav significantly outperforms existing methods in success rate (SR) and Success weighted by Path Length (SPL), offering a lightweight, interpretable, and robust navigation framework without costly online VLM reasoning.

Abstract

Explore similar work

Rethinking Embodied Navigation via Relational Inductive Bias

IntentNav: Learning Spatial-Visual Object Navigation from Human Demonstrations

EffiNav: Fusing Depth and Vision-Language for Efficient Object Goal Navigation