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

Abstract

Autonomous odor source localization remains a challenging problem for aerial robots due to turbulent airflow, sparse and delayed sensory signals, and strict payload and computation 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 that we train in simulation and deploy 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. We incorporate vision 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.

1. Introduction

How does one find what one cannot see, hear or touch? Researchers have likened navigating by scent to “chasing ghosts” due to humans’ lack of high-bandwidth olfactory perception. Olfaction is the most primitive form of perception, yet artificial intelligence systems have predominantly focused on visual, audio, and linguistic data. This oversight largely stems from the scarcity of olfactory data and the absence of standardized benchmarks, which pose significant challenges for developing and evaluating machine learning models in this domain.

In this study, we introduce a model for olfaction-only navigation, but demonstrate how vision can complement performance. We show how these models can be trained via simulation to run at the edge for scent-based navigation to an odor source. Olfactory navigation is largely absent from robotic functions today, and our motivation for this work was to establish methodology that allows robots to, for example, localize the source of a specified chemical compound, inform automated surgery by mapping olfactory cues to visual components, and inspire alternative forms of underwater navigation via chemical plume tracking. We demonstrate our methodology on a real UAV navigating to a chemical compound.

Due to the hysteresis and non-linearity associated with gas dynamics, navigating by scent is inherently difficult. Current olfaction sensors suffer from drift and long sampling times [1,2,3], and the inability to accurately perceive the presence of a chemical compound analogizes to “chasing ghosts”. Here, we detail a confluence of adaptive learning techniques, hardware design, sensor selection, and computation optimizations that make this feasible. We provide a detailed account of our hardware, methodologies, and evaluation metrics to ensure transparency and reproducibility. The Appendix A, Appendix B, Appendix C, Appendix D, Appendix E, Appendix F, Appendix G and Appendix H detail a full bill of materials and code to reproduce our results. Figure 1 shows frontal views of the two configurations we used for the UAV.

Figure 1. The UAV equipped with the olfactory processing unit (OPU) and sensor harnesses for scent-based navigation. (A) shows the metal oxide sensor configuration, and (B) shows the electrochemical sensor configuration.

The contributions of our work are as follows:

• A complete, open-source UAV system for odor source localization, integrating custom olfaction hardware, onboard sensing, simulation, and real-world deployment under payload and computation constraints.
• A learning-based navigation method that we train in simulation and deploy on a real UAV, enabling online odor source localization without prior mapping or external positioning infrastructure.
• A minimal-sensor navigation formulation, demonstrating reliable source-seeking behavior using only a small number of olfaction sensors and a single range sensor, with vision serving optionally as a complementary modality.
• Real-world experimental validation in a large indoor environment, showing consistent source localization behavior and providing qualitative and quantitative insights into plume navigation under realistic airflow conditions.
• A demonstration that our framework generalizes across sensing mediums.

This work emphasizes system design and methodological integration, on which one could expand with statistically powered evaluations across a wider range of environments and conditions.

The code, data, and designs are available at https://github.com/KordelFranceTech/ChasingGhosts (accessed on 26 April 2026).

2. Related Work

Early UAV-based olfaction systems have focused on gas distribution mapping (GDM), where a robot follows a predefined flight pattern to reconstruct a concentration field and estimate a source location post hoc. Burgués et al. [4] presented a representative example, employing a quadrotor to sample gas concentrations along a sweeping trajectory and infer the source location from the reconstructed map. While effective for environmental monitoring, such approaches do not perform online source-seeking behavior and require substantial spatial coverage prior to localization.

Reactive plume-tracing methods draw inspiration from biological odor tracking, enabling robots to move toward an odor source based on local concentration cues. Shigaki et al. [5] demonstrated three-dimensional chemical plume tracing using a quadcopter equipped with dual alcohol sensors and onboard ranging, achieving successful localization in controlled indoor environments. Their work highlights the feasibility of UAV-based plume tracking, but focuses on reactive control strategies rather than learning-based policies trained across environments.

More recently, researchers have explored learning-based and swarm-based approaches to improve robustness in turbulent environments. Duisterhof et al. [6] introduced a multi-agent “Sniffy Bug” system that combines bio-inspired navigation with particle swarm optimization, leveraging inter-agent communication and additional sensing modalities. While effective at the swarm level, such systems rely on greater sensing and coordination infrastructure than a single lightweight UAV typically carries. Research from Zhang et al. [7,8] demonstrates how chemical and visual sensing can work collaboratively on a UAV, but does not demonstrate it in navigation tasks.

In contrast to prior work, this paper focuses on single-agent, learning-based odor source localization under minimal sensing assumptions. Rather than reconstructing gas distributions or relying on swarm coordination, the proposed system learns a navigation policy in a simulation and deploys it directly on a real UAV, enabling online source-seeking behavior in a large indoor environment. The emphasis is on system integration, reproducibility, and sim-to-real transfer, providing a foundation for future statistically powered evaluations and extensions. The works in [4,6] leverage ultrawide-bandwidth (UWB) radio tags placed throughout each course to help the UAV navigate. In contrast, we only allow the UAV to use olfaction (and in the last mile, vision) to guide itself to the odor source, and it has no prior knowledge about the course. Unlike the post-mapping over 160 m² from [4], we achieved real-time localization over 200 m², the largest-sized experimental course so far, without positional aids.

The work by Hassan et al. [9] shows how robots can navigate by olfactory and visual capabilities, but leverages closed-source online large language models to perform the computations as well as the interpretation of the chemical and vision data, methods impractical for real-time edge robotics. In [4,5,6,9], the authors use metal oxide sensors as the olfactory receptor. We demonstrate our framework with metal oxide sensors to show the improvement on previous work, but also demonstrate that our framework generalizes across sensing mediums with the introduction of electrochemical sensors.

Feng et al. constructed a dataset of vision–olfactory relationships in SmellNet [10], but the dataset related exclusively to fruits, nuts and other common foods. Ozguroglu et al. [11] showed how vision and olfaction can link together through contrastive learning, but the scope was exclusive to botany. We built off many of the methods established by [4,6,12]. Namely, we used ethanol as our target compound to maintain parity with their work and provide a level of experimental control.

To our knowledge, this is the first demonstration of stereo-olfactory sensing combined with vision on a UAV over two different chemical sensing mediums. Appendix A presents a concise comparison of our work to related work.

3. Methods

Our primary goal was to demonstrate that UAVs can navigate purely with olfactory sensors to track a target chemical back to its source. However, we recognize that visual sensors can complement olfactory navigation, as seen in several mammals and insects. As a result, we show how understanding the aromas emitted by certain objects can decrease the time needed to localize the target compound through visual confirmation of capturing the odor source.

3.1. Hardware

We desired a solution for olfactory navigation that is financially and technically accessible to robotics and AI researchers while also reaching the precision required to localize a chemical compound. Therefore, we selected off-the-shelf hardware where reasonably possible. Figure 2 shows critical additions to the base UAV airframe, while Appendix H discusses a full bill of materials.

Figure 2. The UAV equipped with the olfactory processing unit (OPU) and sensor harnesses. The figure shows a top-down view of the aircraft: (1) the 2 ballast points needed to balance the aircraft; (2) the battery to power the motherboard required for the olfactory sensors; (3) the wire harnessing leading to the motherboard attached to the belly of the aircraft; (4) the forward-looking time-of-flight sensor for obstacle avoidance; (5) the forward-looking camera; (6) the motherboard and downward-looking time-of-flight infrared sensors attach to the belly, and (7) the two olfactory sensor antennae (MOX sensors shown, but EC sensors lie at the same location when in the proper configuration).

We selected the DJI Tello as the UAV [13], a small open-sourced drone that one can program, but that contains minimal onboard computations. Unlike prior works [4,6] that leverage the Crazyflie UAV platform, we required a slightly larger UAV platform with a camera and higher payload that enables us to carry more sensing capability and computations. Three infrared sensors provide basic stability: two downward-looking sensors for leveling and one forward-looking sensor for obstacle detection. We designed a custom olfactory processing unit (OPU) that contains an ESP-32 microcontroller and a PalmSens EmStat Pico potentiostat and attaches to the bottom of the UAV. The OPU processes the infrared data for obstacle detection and olfaction sensors to inform navigation, and wirelessly streams all data back to a laptop ground station for telemetry. We mechanically modified the Tello UAV to accommodate the harnessing, battery, and circuitry that the OPU requires. This required adding 71 g of ballast to the aft end of the drone in order to counter the forward moment that the additional components create; no additional ballast was needed to adjust for lateral moments. We designed the mechanical brackets and harnessing needed to attach the OPU, olfactory sensors, and infrared sensors to the airframe in SolidWorks [14]. We designed the OPU motherboards using EasyEda [15]. Figure 3 gives a snapshot of the hardware additions while Appendix G and Appendix H present more details on the electrical and mechanical hardware, respectively.

Figure 3. (A) One can additively manufacture the full UAV modification kit for both electrochemical and metal oxide sensing configurations on a single 200 mm × 200 mm print bed. (B) An image of the breakout board we designed to down-select the olfaction sensors for ethanol detection.

3.2. Simulation

We created a digital twin of the modified UAV for two different simulations. We built a physics simulation via SimuLink to model the control loop algorithms for the UAV, test various flight conditions, and establish a control envelope. Figure 4 shows a series of plots from this simulation, establishing the performance of the autopilot control algorithms (for more information on these algorithms, see Appendix B and Table A2 and Table A3). This simulation contains a configuration file that enables extrapolation of our control laws to any other UAV platform with our OPU modification. We built the second simulation using Python 3.12 and the Gymnasium framework [16], which we discuss in detail in the next sections. We leveraged and tuned Kalman filters to smooth signals in the control responses. Taking inspiration from olfactory navigation in insects, we positioned two sensors at the top and fore of the UAV in analogous positions, as the antennae on the silk moth. This differs from previous approaches of olfactory navigation [4,6], where the sensor sat at the top of the body and below the rotors, but it is similar to the locations proposed by [5]. Our positioning allowed us to perform stereo sensing to ascertain the plume direction and avoid volatility from the rotor wash, maximizing the odor signal.

Figure 4. (A) From top to bottom, this panel shows the altitude, velocity, pitch, and roll from the controller response (blue line) and the command (red line) from the control algorithms. (B) A diagram showing a partial flight path of the UAV in the simulation. (C) From top to bottom, this panel shows the olfactory signal responses for ethanol and nitrogen dioxide. Red dotted lines indicate the desired value. Green solid lines indicate the values executed by the autopilot. Blue solid lines indicate the actual achieved values by the UAV.

3.3. Sensors

We desired a generalized solution for olfactory navigation. To aid with down-selection of off-the-shelf olfaction sensors, we designed a small electrical breakout board (see Appendix G). This allowed us to manually test detection algorithms without having to re-configure the mounting hardware and ballast on the drone. The repository associated with this paper contains the circuitry and design of this board. The results of the experiments we performed with this olfactory board showed that metal oxide (MOX) and electrochemical (EC) sensors were the best for tracking ethanol, so we selected them as the final olfactory effectors.

Metal oxide sensors respond rapidly (1–100 Hz) [17] and are typically sensitive to families of molecules, allowing for more general chemical tracking. We selected a pair of SGX Sensortech MICS 6814 sensors (Corcelles-Cormondreche, Switzerland) [18] due to their prevalence in the recent literature [10,17] and the wide array of compounds they detect. These sensors respond to a variety of gases, but we focused on their sensitivity to ethanol for our work here, since ethanol was the target compound.

Metal oxide sensors change their resistance based on their exposure to the target gas. This resistance indicates the gas intensity in the surrounding environment, and one can compute it according to the following equation:

$$R_s=\left[{\displaystyle \frac{V_c}{V_{RL}}}-1\right]·R_L$$

(1)

where $R_S$ indicates the sensor resistance, $V_c$ is the circuit voltage, $V_{RL}$ is the voltage drop over the load resistor, and $R_L$ is the resistance of said load resistor. As Equation (1) is fundamental to the function of the MOX sensor, we designed our optimizations around it. We built off of the methodology from Burgués et al. in [4], in that we designed our algorithm to not depend on absolute concentrations in the parts-per-million or parts-per-billion quantity. Rather, we analyzed the relative change between time steps of $V_{RL}$ (the temporal difference) and used this to inform the UAV on how to move relative to gas measurements.

Electrochemical sensors respond less rapidly, but much more specifically, to individual compounds, allowing for more precise tracking. We leveraged a two-electrode electrochemical sensor from ItalSens designed to maximize diffusivity of compounds over a certain voltage range for chronoamperometry. Chronoamperometry is the process by which one can measure the change in electric current with respect to time over a controlled electrical potential. The Cottrell equation shows this relationship:

$$I={\displaystyle \frac{n_{e}FA{c}_k\sqrt{D_k}}{\sqrt{\pi t}}}$$

(2)

where I denotes the electrical current, measured in amperes; $n_e$ is the number of electrons needed to oxidize one molecule of analyte k; F is the Faraday constant of 96,485 Coulombs per mol; A denotes the planar area of the electrode in square centimeters; $c_k$ defines the initial concentration of the target analyte k in mols per cubic centimeter; $D_k$ defines the diffusion coefficient for analyte k in square centimeters per second; and t is simply the time that the chronoamperometric sequence is running in seconds. Our sensor contains an electrode surface area of 2.25 square centimeters and specifically tracks a target analyte with a reduction potential of 0.8 V according to Fick’s law. We utilized 1-ethyl-3-methylimidazolium tetrafluoroborate ($\left[EMIM\right][B{F}_4$]) as the ionic liquid to act as a transducer over our electrodes.

Typically, a single chronoamperometric sequence operates for anywhere between 6 and 60 s. We borrowed the method from [3] to infer the analyte diffusion in order to speed up the sampling rate to 0.5–1 Hz. Appendix C presents more details on this technique. Unlike our approach with the MOX sensors, we did adjust the fundamental operation of the EC sensors in Equation (2).

3.4. Dataset

We performed pattern recognition over the olfactory data via adaptive temporal difference learning. We did not accumulate a dataset a priori for training the olfactory sensors, as we determined that adaptively learning the plume via real-time exploration approximation was sufficient, following many of the methods that Sutton set forth in [19]. The data for the olfaction–vision models comes from [20], which fuses trusted and peer-reviewed olfactory [21,22,23] and computer vision [24] datasets, and is also the only known olfaction–vision dataset.

While modern reasoning models can associate from which objects an aroma (e.g., lingual descriptors such as “fruity” and “musky”) is coming in an image, others find that they have difficulty in specifying from where a chemical compound (e.g., $C{O}_2$, $C{H}_4$) is coming [10,20,25]. In addition, leveraging a large vision–language reasoning model—most of which are billions or trillions of parameters in size—is neither practical nor necessary for real-time edge robotics. To this end, we constructed our own encoder to model olfactory–vision relationships, discussed in Section 3.5.2.

3.5. Models

3.5.1. Olfaction-Only Navigation

To track a plume, we must make some inference on its direction. We used the time-delay estimation between the left and right sensors to inform the horizontal plume angle of attack against the sensor surface. Let $x_L$ and $x_R$ represent the left and right lateral positions of the sensors onboard the aircraft, respectively. For a planar odor front advecting with velocity u, wind speed s, and sensor separation $\Delta x=x_L-x_R$, we can define the expected sensor time constant $\widehat{\tau }$ as:

$$\widehat{\tau }\approx -{\displaystyle \frac{\Delta x·u}{s^2}}$$

(3)

For a pure lateral baseline $\Delta x$ and known wind speed s, the angle $\varphi$ between the plume travel direction and the longitudinal axis of the aircraft resolves to:

$$\varphi =arcsin{\displaystyle \frac{\widehat{\tau }s}{d}}$$

(4)

This effectively converts a measured time lag into a heading command for the aircraft. We clipped $\varphi$ such that $\varphi \in [-90°,-5°]\cup [5°,90°]$ such that the UAV will hold the current heading for all $\varphi \in (-5°,5°)$. Large angles can be likened to casting (exploratory) behaviors in insects, while smaller angles (meaning that the UAV stays its current course) can be likened to surging (exploitive) behavior.

The UAV samples the readings every second from each of the two MOX sensors and every two seconds for the electrochemical sensors. We applied dual-timescale exponential averaging to detect plume bouts while suppressing sensor drift and turbulence noise. This parallels filtering mechanisms observed in insect olfactory neurons. We calculated this divergence D by subtracting a long exponential moving average (EMA) with period $\beta$ from a short EMA with period $\alpha$, denoted by $E_{\beta }$ and $E_{\alpha }$, respectively:

$$D=E_{\alpha }-E_{\beta }$$

(5)

D can be considered as a temporal-prediction-difference filter, which is mathematically adjacent to TD error computation in reinforcement learning, albeit not a learning mechanism. Although D does not constitute temporal-difference learning in the reinforcement-learning sense, it produces a temporal-prediction error—the difference between fast and slow odor expectations. This is conceptually similar to biological temporal difference filters that insect olfactory pathways use to emphasize odor onset and plume entry events.

We additionally computed a signal line S, which is a smoothed EMA of D with period $\rho$, where $\alpha <\rho <\beta$. This signal line serves as a dynamic expectation of odor-trend behavior. A positive deviation between lines D and S indicates an accelerating odor concentration, consistent with entry into the plume, whereas a negative deviation reflects loss of odor momentum and triggers casting behavior. This dual-timescale architecture parallels both biological olfactory adaptations and momentum-based filtering methods in quantitative finance.

Formally, let $C\left(t\right)$ define a function over time t for the olfaction sensor readings, where $C_L,C_R\in C$ define functions for the left and right sensors, respectively. Then,

$$D\left(t\right)=E_{\alpha }\left(C\left(t\right)\right)-E_{\beta }\left(C\left(t\right)\right)$$

(6)

And the signal line is

$$S\left(t\right)=E_{\rho }\left(D\left(t\right)\right)$$

(7)

While $\varphi$ gives the UAV the incident angle of the plume against the sensors, the bookkeeping of the D and S sequences allows the UAV to determine its proximity relative to the plume. For the experiments shown here, we selected the empirically deduced values of 3, 8, and 5 for $\alpha$, $\beta$, and $\rho$, respectively.

3.5.2. Olfaction–Vision Navigation

For navigation augmented by vision, we trained and evaluated a multimodal machine learning model for understanding olfaction–vision relationships. Figure 5 shows the main architecture. It employs a CLIP-based vision encoder [26] to extract visual features from images. A separate 12-layer fully connected encoder processes chemical data from the olfactory sensors, converting it into 138-dimension olfactory embeddings. We call this model COLIP (contrastive olfaction–language–image pre-training) since it builds on the original CLIP model. A learned linear layer projects both visual and olfactory embeddings into a shared 512-dimension latent space, ensuring compatibility before further processing. A graph associator learns joint relationships between the projected visual and olfactory embeddings. We evaluated two such models to assess tradeoffs in different architectures: a simple fully connected network and a graph attention network (GAT). Both models use the same 12-layer encoder. The GAT captured more complex associations between the projected visual and olfactory embeddings, but we found difficulties in a reliable performance when exporting to a format that could run at the edge on the UAV. As a result, we did not leverage it for the final analysis because the extra computation and inference time required to run the model would contrast with our desire to run all computations fully at the edge. Appendix D presents more details on the training parameters for these models.

Figure 5. Olfaction–vision model architecture. The blue and green boxes construct Component 1 based on COLIP. The yellow boxes construct Component 2. The output from both models informs reasoning for navigation.

Just like its predecessor, COLIP underwent contrastive training using an InfoNCE (information noise contrastive estimation) loss to align olfactory and visual embeddings, encouraging the model to associate corresponding pairs closely in the latent space. These modality pairs encode together, effectively binding them in the process. We optimized these encoders through the InfoNCE loss [27] shown below.

$$L_{I,M}=-log{\displaystyle \frac{exp(q_i^{⊺}{k}_i/\tau )}{exp(q_i^{⊺}{k}_i/\tau )+{\sum }_{j\ne i}exp(q_i^{⊺}{k}_j/\tau )}}$$

(8)

where $\tau$ denotes a scalar temperature that controls the smoothness of the softmax distribution and j defines “negatives”, or unrelated observations. We presumed each example $j\ne i$ in the mini-batch to be a negative. The loss makes the embeddings $q_i$ and $k_i$ closer in Euclidean distance in the joint embedding space. This consequently aligns I with M conveniently for joint learning. A binary classification loss from the graph head learns relationships between modalities, distinguishing between matching and non-matching olfaction–vision pairs.

Many approaches to robotic control struggle with a core tradeoff: the backbones of large vision models are highly general, but too slow for real-time operation, while traditional robot visuomotor policies are fast, but lack generalization [28]. We resolved this tradeoff by introducing a dual-system architecture. Component 1 is the OVM that runs at 0.1–0.2 Hz. It handles high-level scene understanding and multimodal comprehension, enabling broad generalization across objects, environments, and goals. Component 2 is a high-speed reactive vision–olfaction module containing a fine-tuned YOLOv11 [29] for rapid vision inference, running inference at 1–2 Hz. Component 1 mainly confirms whether the acceptance criteria for locating the odor source have been achieved (see Section 3.8 for more on this). Component 2 synchronizes at the same sampling rate as the olfaction sensors, and helps convert the latent representations from the OVM into precise, continuous UAV actions.

This decoupled design lets each component function at its optimal temporal scale. Component 2 can deliberate slowly over abstract goals, while Component 1 executes rapid, low-latency commands in response to dynamic conditions. This relieves the need to model increased complexity in our plume models and simulations. Furthermore, the separation of both components allows for independent development and iterations.

3.6. Navigation and Plume Models

Plume models are inherently complex and require advanced computational fluid dynamics engines to achieve reasonable approximations. We sought a set of algorithms that generalize well to all chemical compounds and follow simple rules based on environmental evidence, not unlike those that insects follow [12,30,31]. The primary navigation model was a simple bout detection algorithm [6,32] in which the first and second derivatives align with the short and long moving averages of the olfactory signal coupled with inertial data to determine the UAV actions. For brevity, we denote this model as olfactory inertial odometry (OIO), building off the principles from [33]. We modeled plumes with simple Gaussian time-series processes, inserting blanks [34] into the data to resemble wind shifts and plume volatility. The robot has seven actions available depending on temporal plume dynamics: surge forward, cast in one of four directions, pause, or land. Selecting the “land” action indicates that the UAV is confident that it has found the plume source. Our perception reasoning is highly susceptible to hysteresis due to the ability of shifting plume dynamics. We employed tuned Kalman filters and sensor synchronization to counter this. We implemented a custom plume environment using the Gymnasium framework [16] for additional simulations in plume tracking. We designed factors such as the temperature, relative humidity, barometric pressure, wind direction, wind magnitude, wind sparsity, and air chemical composition to all be tunable hyperparameters of the plume. The design follows Gaussian plume principles [31] and a Dryden turbulence model [35]. Appendix F presents more details about this plume environment and its construction.

We implemented a second navigation model based on reinforcement learning (RL) to provide a more comprehensive assessment of the navigation. Q-learning [36] is a standard temporal difference RL algorithm that many applications use. Due to the dynamism of plumes and the UAV’s frequent probability to travel off plume, we employed eligibility traces to harmonize immediate and long-term rewards. $TD\left(\lambda \right)$ algorithms leverage eligibility traces to give credit to more recent $\lambda$ states instead of only the current state ($\lambda =0$). Consequently, we leveraged the TD(λ) derivatives of Q-learning and Expected SARSA, called $Q\left(\lambda \right)$ and Expected SARSA(λ), respectively [37]. While we trained both models in the simulation, we found Expected SARSA $\left(\lambda \right)$ to be the model most robust to reasoning about blank pockets of air, and thus selected it as a finalist for evaluation in our experiments. Expected SARSA$\left(\lambda \right)$ proved most robust for air with continuous gradients, but this scenario is not practical for real-world applications. Appendix E presents the exact training parameters for these models.

With either model, we discretized the aircraft’s movements such that it can only move 10 cm/s or ${\displaystyle \frac{\pi }{2}}$ radians/s at each time step. In essence, we gave identical linear and angular acceleration commands for every command. We followed a policy that allows the UAV to move and wait for 2 s to sample continually until done. For each “move” command, the magnitudes of acceleration and velocity were identical to provide a level of experimental control and reproducibility. The consistency here also ensured that we did not encounter further nonlinearities induced by the payload bias. We found that making the actions continuous was not necessary, as doing so added more complexity to our simulations and did not substantially improve them. In addition, the infrared sensors enabled the UAV to make a decision on how to move forward. We acknowledge that enabling continuous action spaces could present an opportunity for future work.

3.7. Course

Finally, we transferred our solution to the real world with flight tests on the UAV. We deployed our UAV over a 200 m² course, depicted in Figure 6. We started by placing 100 mL of ethanol within a home diffuser in Room 2, as the checkered flag indicates. The diffuser diffused the ethanol at a rate of 2.16 mL per minute. We placed the diffuser near an 80 cm home fan that exhausted air at 1415 L per minute in order to develop a small plume for traceability of the compound. The robot started at the far adjacent end of the map, as the UAV icon in the figure indicates. We observed 5 runs each with the target compound lying in both Rooms 1 and 2. At the beginning of each run, we gave the compound 5 min to disperse and form the start of a plume before we commanded the UAV for takeoff. After takeoff was initiated, no humans intervened with the UAV until it sent a response back to the ground station that it had found the source of the target compound. Between each run, we vented the course and took air measurements to ensure that the next iteration would begin with a clean baseline and to encourage as reproducible results as possible. Due to the length of time needed to properly vent the course, the uncontrollable HVAC schedules within the test building, and other experimental controls needed to re-initialize to a stable environment, the number of statistical runs we could perform was limited. As a result, for each of the four tasks, we analyzed five trials. We note that prior work from [4,6] had a low count of statistical runs (identical to ours), which we suspect these factors also caused.

Figure 6. Illustration of the course developed for olfactory navigation by the UAV. (A) shows a bird’s-eye view of the course, where the ground station (laptop icon) is located at the top left and the UAV (quadcopter icon) takes off next to it; Room 2 hosts the target compound and the plume is formed from here. (B) shows an isometric view of the ideal path that the UAV should pursue to locate the source. The blue dotted line indicates the minimum optimal path that the UAV should travel to localize the odor source, though this was never achieved.

3.8. Task Termination Criteria

One defining contribution of our work is how a UAV decides that it is “done”. For the robot to determine whether it has found the source of the gas, it must evaluate a hypothesis about whether it believes it has found the highest concentration of the plume given its observed samples. On our selected UAV, the IMU bias from the added payload was significant, making long-term positional estimation very difficult. The UAV could perform all of the maneuvers commanded, but the added payload dampened the magnitude of each command. We found this dampening factor non-linear and partially dependent on battery life. This made the error of the commands difficult to model, but we found short-term estimation more practical to filter via extended Kalman filters and reasoned via OIO accordingly. See Figure 7 as an example of the IMU replay that occurred over the map detailed in Figure 6.

We concluded that the UAV reached the gas source when it landed within 1 m of the diffuser. To evaluate when the UAV reached the gas source, we constructed a minimum-variance unbiased point estimator $\widehat{C}$ to approximate the true highest concentration of the gas within the environment, given by:

$$\widehat{C}=m{\displaystyle \frac{k}{k+1}}-1$$

(9)

where m is the largest concentration that the olfaction sensors observed and k is the total number of samples observed. We assumed that, once a chemical packet has been observed, it can be observable again, as in drawing with replacement [38,39].

Figure 7. The IMU replay of the UAV traveling over the whole course is shown here. Unfiltered olfactory inertial odometry (OIO) [33,40] and filtered OIO via EKF are shown alongside naked measurements from the IMU. The addition of the olfaction sensors and OPU as payload significantly biased the IMU. This made compound mapping via long-term inertial odometry nearly impossible and drove our decision for the task termination criteria in Section 3.8. Measurements are in centimeters.

Figure 7. The IMU replay of the UAV traveling over the whole course is shown here. Unfiltered olfactory inertial odometry (OIO) [33,40] and filtered OIO via EKF are shown alongside naked measurements from the IMU. The addition of the olfaction sensors and OPU as payload significantly biased the IMU. This made compound mapping via long-term inertial odometry nearly impossible and drove our decision for the task termination criteria in Section 3.8. Measurements are in centimeters.

The point estimate becomes reliable only after long sessions with several observed samples. However, because we desired a generalized solution to both short and long exploration sessions, we computed a confidence interval for $\widehat{C}$. We computed the p-th and q-th quantiles of the sample maximum m to form the interval $[\widehat{C}{q}^{1/k},\widehat{C}{p}^{1/k}$], which yielded the corresponding confidence interval for the maximum estimated gas source concentration over the full population of gas samples:

$$\left[{\displaystyle \frac{m}{q^{1/k}}},{\displaystyle \frac{m}{p^{1/k}}}\right]$$

(10)

As an example, assume that the UAV has observed 20 samples and believes that it has found the source of the gas with an observed final concentration m. The point estimate for $k=20$ samples would be $1.05m$ based on Equation (9). By Equation (10), taking a 95% confidence interval would return

$$[{\displaystyle \frac{m}{{0.975}^{1/20}}},{\displaystyle \frac{m}{{0.025}^{1/20}}}]=[1.001m,1.2025m]\approx [m,1.2m]$$

(11)

This indicates that, given the observed plume concentrations, we can be 95% confident that the true highest concentration of the plume lies between the current observed highest concentration m and a concentration 20% larger. From here, the UAV can determine whether to keep navigating or to end. In general, we strive to stay within 25% of the expected maximum odor concentration based on a 95% confidence interval, so in this case, the UAV would land and stop navigating.

4. Results

Table 1 shows the final results for our experiments. For all trials, the UAV successfully found the source of the target compound, which was our minimum bar for success. We observed that electrochemical-based olfactory sensors require more time to complete the task in general, and we suspect that this is due to the fact that EC sensors are more sensitive to perturbations in the target compound’s concentration, which could cause them to overfit to noise or small eddy currents in the plume. For both MOX and EC sensors, we noted a small bump in performance from adding vision as a modality by which to localize the odor. In all tasks, the UAV only used olfaction to navigate. OIO showed more consistency in navigation times than RL. We suspect that this is due to RL over-optimizing the navigation path by attempting to predict ahead or weighing states too far in the past for future decisions due to the eligibility traces. In future work, we expect that wrapping OIO with RL will become more competitive (or objectively better) than naked OIO with further tuning. In the olfaction + vision tasks, the UAV only activated the camera when it could no longer maximize the olfactory gradient, which usually occurred when the UAV was confident that it had reached the target compound based on the termination criteria. When the UAV used only olfaction, we observed that it would spend several seconds slowly moving around the chemical source, trying to maximize the olfactory signal. Because diffusion of the gas is more uniform the closer one is to the gas’s source, an apparent gradient may not be easy to determine, and the UAV will waste several seconds moving around very close to the source. This is where vision became additive: using a visual confirmation that the UAV has found the source allows the UAV to not waste time and battery chasing small gradients. As a consequence, vision shortened the navigation times by shaving a few seconds off the end of each trial, where olfaction-only navigation would have kept the UAV flying.

Table 1. Final results.

Method	Sensor Type	Nav. Algorithm	Average Time ${\mathit{\mu }}_{\mathit{t}}$ (s)	Sigma Time ${\mathit{\sigma }}_{\mathit{t}}$ (s)	Best Time ${\mathit{\beta }}_{\mathit{t}}$ (s)
Olfaction	MOX	OIO	98.38	14.84	71.59
Olfaction	EC	OIO	112.22	16.70	97.65
Olfaction	MOX	E.SARSA$\left(\lambda \right)$	103.99	14.89	82.73
Olfaction	EC	E.SARSA$\left(\lambda \right)$	121.85	16.77	97.10
Olfaction + Vision	MOX	OIO	94.03	14.68	66.12
Olfaction + Vision	EC	OIO	107.71	17.07	80.50

5. Discussion and Limitations

The most important conclusion we drew from our experiments was that olfaction is enough to navigate when localizing to an odor source. We believe that this is the first documented attempt to enable a robot to navigate purely by scent without any planning, a priori knowledge of the environment, or location aids such as UWB radio tags. We attempted to show that our results generalize by performing experiments with two different olfaction sensor types. However, we acknowledge that a larger variety of chemical sensors exists. In addition, we performed our work here over only one chemical compound: ethanol. This was intentional in order to build off of the assumptions and work by Burgues et al. [4] and Duisterhof et al. [6], as they demonstrated olfactory navigation with radio tags and metal oxide sensors in tracking ethanol. With our work here, we attempted to establish a generalizable framework that builds off the accepted state of the art, and that can enable intuitive extrapolation of our hardware and algorithms to other compounds with little or no tuning. We are performing ongoing work to support this and demonstrate said framework on a variety of gases in outdoor conditions and wider kinematic envelopes.

We note that our results show that the UAV localized itself to the chemical source on every trial, while the work of Shigaki et al. [5] noted a non-perfect localization rate. We are also performing ongoing work to demonstrate the generalization of our framework over other environments, chemical compounds, and sensor types such as graphene-based sensors, spectroscopy sensors, and photo-ionization detectors. We acknowledge that exact replication of any olfactory navigation task is difficult due to the inherent non-linearities of gas dynamics and the difficulty in replicating exact initial chemistry conditions. As a result, the focus of our work here was to establish a simple, reproducible, and generalized framework on which other researchers can build to conduct additional experiments that will lead to further improvements and an increased reproducibility.

6. Conclusions

Scent-based navigation is a very nuanced robotics task. Organisms that leverage olfaction to perceive the world optimize over various criteria, yet none of them are standardized. With our framework and experiments here, we show that, to localize a UAV to a chemical source, olfactory perception is enough. We incorporated many nature-inspired principles to define methods for the first instance of stereo olfaction–vision odor localization and demonstrate their practicality with real aerial robots. We make our code, simulations, hardware schematics, and datasets available to the community to encourage advancement in the field. We hope that the techniques presented here encourage more work in olfactory robotics and bring us one step closer to enabling the sense of smell for machines.