OdorSpaces: Visual-Olfactory Embodied Navigation in 3D Environments

Changyeon Lee; Kelin Yu; Zheng Wei; Shuna Ni; Ruohan Gao

cs.RONov 3, 2026

OdorSpaces: Visual-Olfactory Embodied Navigation in 3D Environments

Authors: Changyeon Lee, Kelin Yu, Zheng Wei, Shuna Ni, Ruohan Gao

Organizations: University of Maryland, College Park

Abstract

Olfaction provides critical information about airborne chemical cues, yet it remains largely underexplored in embodied AI, with limited prior machine olfaction work focusing primarily on semantic odor perception. We introduce OdorSpaces, a physics-based simulation platform and benchmark for visual-olfactory embodied navigation. OdorSpaces augments Habitat-Sim with 3D odor transport and dispersion simulated by Fire Dynamics Simulator (FDS), enabling agents to navigate photorealistic indoor environments using both visual and olfactory observations. The benchmark includes two representative source-seeking scenarios: invisible gas-leak localization and early-stage fire localization. We develop an initial multimodal reinforcement learning policy that integrates visual observations and olfactory measurements over time. Experiments show that the policy outperforms a diverse set of baselines, establishing a strong reference for future research on visual-olfactory embodied navigation.

Figures & tables

Figure 1: OdorSpaces is a visual-olfactory navigation benchmark where embodied agents navigate toward non-visible odor sources using vision, odor, and pose cues. It uses Fire Dynamics Simulator (FDS) to model physics-based 3D odor fields within Habitat-Sim environments, covering two representative source-seeking scenarios: propane gas leak and methanol pool fire.

Figure 2: Example of the FDS simulation setup.

Figure 3: Visual-olfactory navigation policy model architecture. We use a multimodal policy that encodes three observation streams: visual observation $V_{t}$ , olfactory observation $O_{t}$ , and pose-history observation $P_{t}$ . The encoded features are fused and passed to a recurrent memory backbone, whose output $z_{t}$ is used by the policy and value heads for action prediction and value estimation.

Method	Replica		Matterport3D
Method	SPL $(\%,\uparrow)$	Final Dist. $(\mathrm{m},\downarrow)$	SPL $(\%,\uparrow)$	Final Dist. $(\mathrm{m},\downarrow)$
Temporal Gradient Search	$4.59\pm 0.44$	$6.28\pm 0.16$	$1.83\pm 0.96$	$7.79\pm 0.11$
Biomimetic Search (Moth)	$36.11\pm 0.51$	$4.61\pm 0.25$	$33.56\pm 1.69$	$4.14\pm 0.37$
Chemotactic Search	$17.33\pm 2.50$	$5.45\pm 0.22$	$27.10\pm 2.70$	$4.70\pm 0.39$
Dual-Recurrent Threshold RL	$54.50\pm 5.50$	$2.03\pm 1.03$	${\mathbf{65.41\pm 5.15}}$	$\mathbf{1.54\pm 0.38}$
Ours	$\mathbf{65.26\pm 6.59}$	$\mathbf{1.48\pm 0.47}$	$63.82\pm 7.85$	${1.59\pm 0.67}$

Table 1: Propane Gas Leak Evaluation Results. SPL in percent; final distance in meters.

Method	Replica		Matterport3D
Method	SPL $(\%,\uparrow)$	Final Dist. $(\mathrm{m},\downarrow)$	SPL $(\%,\uparrow)$	Final Dist. $(\mathrm{m},\downarrow)$
Temporal Gradient Search	$0.00\pm 0.00$	$7.27\pm 0.15$	$0.00\pm 0.00$	$8.73\pm 0.13$
Biomimetic Search (Moth)	$0.00\pm 0.00$	$7.32\pm 0.24$	$0.00\pm 0.00$	$8.75\pm 0.04$
Chemotactic Search	$0.00\pm 0.00$	$7.25\pm 0.04$	$0.00\pm 0.00$	$8.89\pm 0.03$
Dual-Recurrent Threshold RL	$17.58\pm 3.45$	$5.07\pm 0.49$	$2.74\pm 1.83$	$7.60\pm 0.13$
Ours	$\mathbf{24.72\pm 3.69}$	$\mathbf{4.62\pm 0.41}$	$\mathbf{14.16\pm 2.96}$	$\mathbf{4.57\pm 0.65}$

Table 2: Methanol Pool Fire Evaluation Results. SPL in percent; final distance in meters.

Figure 4: Navigation trajectory examples. The first row shows propane gas leak episodes, and the second row shows methanol pool fire episodes. The first two columns of trajectory examples are success trajectories, where the last column shows representative failure cases.

Method	SPL (%) $\uparrow$	Final Dist. (m) $\downarrow$
O	$14.55\pm 3.53$	$5.05\pm 0.20$
O + V	$19.14\pm 2.78$	$4.80\pm 0.25$
O + P	$16.49\pm 3.34$	$5.22\pm 0.24$
O + V + P	$\mathbf{24.72\pm 3.69}$	$\mathbf{4.62\pm 0.41}$

Table 3: Input modality ablation. We compare ablated variants of our policy using odor (O), vision (V), pose (P), and their combinations as inputs for methanol pool fire in Replica.

Ablation	Method	SPL (%) $\uparrow$	Final Dist. (m) $\downarrow$
Input	O	$49.47\pm 2.99$	$2.88\pm 0.29$
	O + V	$58.31\pm 3.18$	$2.52\pm 0.53$
	O + P	${60.64\pm 5.84}$	${1.98\pm 0.24}$
	O + V + P	$\mathbf{65.26\pm 6.59}$	$\mathbf{1.48\pm 0.47}$
Curriculum	Without Curriculum	${53.57\pm 5.49}$	${2.54\pm 0.31}$
Curriculum	With Curriculum	$\mathbf{65.26\pm 6.59}$	$\mathbf{1.48\pm 0.47}$

Table 4: Propane gas leak ablations. We evaluate the effects of input modalities (odor (O), vision (V), and pose history (P)), distance-based curriculum, reward components, temporal and spatial olfactory sensing, and recurrent memory in the Replica propane gas leak scenario.

Ablation	Method	SPL (%) $\uparrow$	Final Dist. (m) $\downarrow$
Curriculum	Without Curriculum	$20.43\pm 6.73$	$4.85\pm 0.50$
Curriculum	With Curriculum	$\mathbf{24.72\pm 3.69}$	$\mathbf{4.62\pm 0.41}$
Reward	Stop Reward Only	$0.00\pm 0.00$	$6.17\pm 0.49$
Reward	Stop Reward + Distance Reward	$\mathbf{24.72\pm 3.69}$	$\mathbf{4.62\pm 0.41}$
Temporal	Single Timestep	$21.95\pm 0.49$	$\mathbf{4.33\pm 0.29}$
Temporal	Multiple Timesteps	$\mathbf{24.72\pm 3.69}$	${4.62\pm 0.41}$

Table 5: Methanol pool fire ablations. We evaluate the effects of distance-based curriculum, reward components, temporal and spatial olfactory sensing, and recurrent memory in the Replica methanol pool fire scenario.

	Test Distance (m)
Train Dist.	5 m	7 m	9 m	11 m	13 m	15 m	20 m
7 m	${57.26\pm 5.86}$	$63.00\pm 0.34$	$58.77\pm 8.80$	$45.13\pm 5.23$	$36.52\pm 5.77$	$29.40\pm 8.02$	$23.61\pm 4.12$
9 m	$60.86\pm 1.55$	$\mathbf{72.80\pm 4.18}$	${74.37\pm 2.05}$	$76.79\pm 2.66$	$57.66\pm 2.85$	$49.37\pm 3.17$	$44.90\pm 3.22$
11 m	$\mathbf{63.65\pm 4.24}$	$71.49\pm 2.76$	$\mathbf{80.77\pm 1.94}$	$\mathbf{79.60\pm 7.11}$	$\mathbf{79.76\pm 6.81}$	$\mathbf{66.77\pm 5.72}$	$\mathbf{49.17\pm 4.42}$

Table 6: Generalization to longer spawn distances. SPL (%) for different training and test spawn distances in the Replica apartment_0 methanol pool fire scenario.

Odor Source Localization

See all 8 papers

Feb 23, 2026

Chasing Ghosts: A Simulation-to-Real Olfactory Navigation Stack with Optional Vision Augmentation

Autonomous odor source localization remains a challenging problem for aerial robots due to turbulent airflow, sparse and delayed sensory signals, and strict payload and compute constraints. While prior unmanned aerial vehicle (UAV)-based olfaction systems have demonstrated gas distribution mapping or reactive plume tracing, they rely on predefined coverage patterns, external infrastructure, or extensive sensing and coordination. In this work, we present a complete, open-source UAV system for online odor source localization using a minimal sensor suite. The system integrates custom olfaction hardware, onboard sensing, and a learning-based navigation policy trained in simulation and deployed on a real quadrotor. Through our minimal framework, the UAV is able to navigate directly toward an odor source without constructing an explicit gas distribution map or relying on external positioning systems. Vision is incorporated as an optional complementary modality to accelerate navigation under certain conditions. We validate the proposed system through real-world flight experiments in a large indoor environment using an ethanol source, demonstrating consistent source-finding behavior under realistic airflow conditions. The primary contribution of this work is a reproducible system and methodological framework for UAV-based olfactory navigation and source finding under minimal sensing assumptions. We elaborate on our hardware design and open source our UAV firmware, simulation code, olfaction-vision dataset, and circuit board to the community. Code, data, and designs will be made available at https://github.com/KordelFranceTech/ChasingGhosts.

Kordel K. France, Ovidiu Daescu, Latifur Khan +1

Dept. of Computer Science University of Texas at Dallas Richardson, TX, USA

Odor Source Localization Robot Navigation UAV Navigation

Ask PDF

Sep 20, 2026

OmniEcho: Audio-Visual Spatial Understanding for Omni-Modal Embodied Agents

Humans can effortlessly localize the direction of a sound source and integrate it with visual cues for reasoning, yet this remains challenging for embodied agents. In particular, it is still unclear how to effectively evaluate and model spatial audio understanding in embodied settings. To address this gap, we introduce OmniEchoBench, a unified benchmark for spatial audio-visual perception and audio-vision-language navigation. OmniEchoBench comprises six tasks over 197 real-world spatial audio-visual scenes, 2,972 question-answer pairs, and 900 navigation samples with first-order ambisonics (FOA) audio collected from 30 real-world environments. To enable scalable training supervision, we develop a controllable rendering pipeline for spatial audio. It preserves geometric consistency among sound sources, visual observations, and agent trajectories. Building on this, we propose OmniEcho, a spatially aware omni-modal model. It introduces an FOA spatial encoder alongside a pretrained semantic audio pathway. Extensive experiments show that OmniEcho achieves state-of-the-art performance on spatial audio-visual perception. For our sound-guided navigation, OmniEcho reaches a performance level close to that of traditional vision-language navigation. These results demonstrate that spatial audio can serve as a valuable signal for embodied scene reasoning and navigation, while also highlighting fine-grained spatial localization and distance estimation as important open challenges. Our code and data will be available in https://github.com/PKU-VaLuE-Lab/OmniEcho/tree/main

Ruixun Liu, Yuxuan Wang, Jiacheng Xie +10

School of Intelligence Science and Technology, Peking University · State Key Laboratory of General Artificial Intelligence, Peking University · Alibaba Token Hub, Alibaba Group +1

Audio-Visual Understanding Spatial Reasoning Benchmarks Embodied Navigation Audio-Visual Navigation Spatial Audio

Ask PDF

Jun 25, 2026

See & Sniff: Learning Visuo-Olfactory Representations

While modern multimodal models integrate vision with language, audio, or touch, olfaction remains largely unexplored due to the lack of paired visuo-olfactory data. We introduce SmellNet-V, a scalable visuo-olfactory dataset built on the insight that odor identity is largely invariant to visual transformations within a semantic category. This allows us to synthetically pair smell-only samples with semantically aligned in-the-wild web images, converting a unimodal olfactory dataset into a cross-modal benchmark without costly co-collection. Building on this dataset, we propose See & Sniff, a self-supervised framework that learns joint visuo-olfactory representations via dense local alignment and naturally produces smell saliency maps for spatial grounding of odor sources. We further introduce pixel-level smell localization task and a benchmark for evaluation. Our method surpasses smell-only baselines by 7% in smell classification from smell alone and generalizes to cross-modal retrieval and smell localization, establishing visuo-olfactory learning as a new direction in multimodal perception.

Seongyu Kim, Seungwoo Lee, Hyeonggon Ryu +2

Korea Advanced Institute of Science and Technology, Korea · Hankuk University of Foreign Studies, Korea · Ulsan National Institute of Science and Technology, Korea

Representation Learning Cross-Modal Representation Learning Cross-Modal Alignment Multimodal Understanding

Ask PDF

Abstract

Odor Source Localization

Explore similar work

Chasing Ghosts: A Simulation-to-Real Olfactory Navigation Stack with Optional Vision Augmentation

OmniEcho: Audio-Visual Spatial Understanding for Omni-Modal Embodied Agents

See & Sniff: Learning Visuo-Olfactory Representations