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

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

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 VtV_{t} , olfactory observation OtO_{t} , and pose-history observation PtP_{t} . The encoded features are fused and passed to a recurrent memory backbone, whose output ztz_{t} is used by the policy and value heads for action prediction and value estimation.
MethodReplicaMatterport3D
SPL (%,↑)(\%,\uparrow)Final Dist. (m,↓)(\mathrm{m},\downarrow)SPL (%,↑)(\%,\uparrow)Final Dist. (m,↓)(\mathrm{m},\downarrow)
Temporal Gradient Search4.59±0.444.59\pm 0.446.28±0.166.28\pm 0.161.83±0.961.83\pm 0.967.79±0.117.79\pm 0.11
Biomimetic Search (Moth)36.11±0.5136.11\pm 0.514.61±0.254.61\pm 0.2533.56±1.6933.56\pm 1.694.14±0.374.14\pm 0.37
Chemotactic Search17.33±2.5017.33\pm 2.505.45±0.225.45\pm 0.2227.10±2.7027.10\pm 2.704.70±0.394.70\pm 0.39
Dual-Recurrent Threshold RL54.50±5.5054.50\pm 5.502.03±1.032.03\pm 1.0365.41±5.15{\mathbf{65.41\pm 5.15}}1.54±0.38\mathbf{1.54\pm 0.38}
Ours65.26±6.59\mathbf{65.26\pm 6.59}1.48±0.47\mathbf{1.48\pm 0.47}63.82±7.8563.82\pm 7.851.59±0.67{1.59\pm 0.67}
Table 1: Propane Gas Leak Evaluation Results. SPL in percent; final distance in meters.
MethodReplicaMatterport3D
SPL (%,↑)(\%,\uparrow)Final Dist. (m,↓)(\mathrm{m},\downarrow)SPL (%,↑)(\%,\uparrow)Final Dist. (m,↓)(\mathrm{m},\downarrow)
Temporal Gradient Search0.00±0.000.00\pm 0.007.27±0.157.27\pm 0.150.00±0.000.00\pm 0.008.73±0.138.73\pm 0.13
Biomimetic Search (Moth)0.00±0.000.00\pm 0.007.32±0.247.32\pm 0.240.00±0.000.00\pm 0.008.75±0.048.75\pm 0.04
Chemotactic Search0.00±0.000.00\pm 0.007.25±0.047.25\pm 0.040.00±0.000.00\pm 0.008.89±0.038.89\pm 0.03
Dual-Recurrent Threshold RL17.58±3.4517.58\pm 3.455.07±0.495.07\pm 0.492.74±1.832.74\pm 1.837.60±0.137.60\pm 0.13
Ours24.72±3.69\mathbf{24.72\pm 3.69}4.62±0.41\mathbf{4.62\pm 0.41}14.16±2.96\mathbf{14.16\pm 2.96}4.57±0.65\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.
MethodSPL (%) ↑\uparrowFinal Dist. (m) ↓\downarrow
O14.55±3.5314.55\pm 3.535.05±0.205.05\pm 0.20
O + V19.14±2.7819.14\pm 2.784.80±0.254.80\pm 0.25
O + P16.49±3.3416.49\pm 3.345.22±0.245.22\pm 0.24
O + V + P24.72±3.69\mathbf{24.72\pm 3.69}4.62±0.41\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.
AblationMethodSPL (%) ↑\uparrowFinal Dist. (m) ↓\downarrow
InputO49.47±2.9949.47\pm 2.992.88±0.292.88\pm 0.29
O + V58.31±3.1858.31\pm 3.182.52±0.532.52\pm 0.53
O + P60.64±5.84{60.64\pm 5.84}1.98±0.24{1.98\pm 0.24}
O + V + P65.26±6.59\mathbf{65.26\pm 6.59}1.48±0.47\mathbf{1.48\pm 0.47}
CurriculumWithout Curriculum53.57±5.49{53.57\pm 5.49}2.54±0.31{2.54\pm 0.31}
With Curriculum65.26±6.59\mathbf{65.26\pm 6.59}1.48±0.47\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.
AblationMethodSPL (%) ↑\uparrowFinal Dist. (m) ↓\downarrow
CurriculumWithout Curriculum20.43±6.7320.43\pm 6.734.85±0.504.85\pm 0.50
With Curriculum24.72±3.69\mathbf{24.72\pm 3.69}4.62±0.41\mathbf{4.62\pm 0.41}
RewardStop Reward Only0.00±0.000.00\pm 0.006.17±0.496.17\pm 0.49
Stop Reward + Distance Reward24.72±3.69\mathbf{24.72\pm 3.69}4.62±0.41\mathbf{4.62\pm 0.41}
TemporalSingle Timestep21.95±0.4921.95\pm 0.494.33±0.29\mathbf{4.33\pm 0.29}
Multiple Timesteps24.72±3.69\mathbf{24.72\pm 3.69}4.62±0.41{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 m7 m9 m11 m13 m15 m20 m
7 m57.26±5.86{57.26\pm 5.86}63.00±0.3463.00\pm 0.3458.77±8.8058.77\pm 8.8045.13±5.2345.13\pm 5.2336.52±5.7736.52\pm 5.7729.40±8.0229.40\pm 8.0223.61±4.1223.61\pm 4.12
9 m60.86±1.5560.86\pm 1.5572.80±4.18\mathbf{72.80\pm 4.18}74.37±2.05{74.37\pm 2.05}76.79±2.6676.79\pm 2.6657.66±2.8557.66\pm 2.8549.37±3.1749.37\pm 3.1744.90±3.2244.90\pm 3.22
11 m63.65±4.24\mathbf{63.65\pm 4.24}71.49±2.7671.49\pm 2.7680.77±1.94\mathbf{80.77\pm 1.94}79.60±7.11\mathbf{79.60\pm 7.11}79.76±6.81\mathbf{79.76\pm 6.81}66.77±5.72\mathbf{66.77\pm 5.72}49.17±4.42\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.
More in this topic

Odor Source Localization

See all 8 papers

Explore similar work

CardsList

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

    Feb 23, 2026Kordel K. France, Ovidiu Daescu, Latifur Khan +1Odor Source LocalizationRobot Navigation

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

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

    Sep 20, 2026Ruixun Liu, Yuxuan Wang, Jiacheng Xie +10Audio-Visual UnderstandingSpatial Reasoning Benchmarks

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

  3. See & Sniff: Learning Visuo-Olfactory Representations

    Jun 25, 2026Seongyu Kim, Seungwoo Lee, Hyeonggon Ryu +2Representation LearningCross-Modal Representation Learning

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