💻 computer science

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

This paper presents a reproducible, open-source framework for autonomous UAV odor source localization that utilizes a minimal sensor suite and a simulation-to-real learning-based navigation policy to find odor sources in turbulent environments without requiring explicit gas mapping or external infrastructure, while optionally incorporating vision to enhance performance.

Original authors: Kordel K. France, Ovidiu Daescu, Latifur Khan, Rohith Peddi

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kordel K. France, Ovidiu Daescu, Latifur Khan, Rohith Peddi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are in a pitch-black room, and someone has hidden a bottle of perfume. You can't see it, you can't hear it, and you can't touch it. The only clue you have is the faint, swirling scent drifting through the air. How do you find it?

This is the challenge of olfactory navigation (finding things by smell). For robots, this is like "chasing ghosts" because air currents are chaotic, smells are patchy, and sensors are often slow or unreliable.

This paper, titled "Chasing Ghosts," describes a team of researchers who taught a small drone to find a hidden smell source without using GPS, maps, or even a camera (unless it really needed one). Here is the story of how they did it, broken down into simple concepts.

1. The Robot: A Drone with a "Super-Nose"

The researchers took a standard, off-the-shelf toy drone (a DJI Tello) and gave it a custom "sniffer" backpack.

The Hardware: They attached two types of "noses" to the drone:
- Metal Oxide Sensors (MOX): These are like a dog's nose. They react very fast to a wide variety of smells but aren't great at telling exactly what the smell is.
- Electrochemical Sensors (EC): These are like a specialized bloodhound. They are slower but can identify specific chemicals (like ethanol/alcohol) with high precision.
The Setup: They placed these sensors on the front of the drone, sticking out like antennae on a moth. This allows the drone to smell the air on its left and right sides separately, helping it figure out which way the wind is blowing the scent.

2. The Training Ground: A Virtual Wind Tunnel

Before letting the drone fly in the real world, the team built a digital twin (a video game version) of their drone and the environment.

The Problem: Real wind is messy. Smells don't flow in straight lines; they break into swirls and pockets of clean air.
The Solution: They programmed the drone to learn in this simulation. They taught it a simple rule: "If the smell gets stronger, keep going. If it gets weaker, turn around and cast a net (search in a zigzag)."
The Learning: They used a type of AI called Reinforcement Learning. Think of it like training a dog with treats. Every time the drone moved closer to the smell, it got a "virtual treat." Every time it got lost, it got a "virtual scolding." Eventually, the drone learned the best way to chase the scent.

3. The Strategy: "Surge and Cast"

The drone doesn't just fly in a straight line. It mimics how insects (like moths) find mates:

Surging: When the drone smells the target strongly, it surges forward confidently.
Casting: When the smell disappears (because the wind shifted), the drone stops and swings back and forth (casts) to catch the scent again.
Stereo Smelling: By comparing the left and right sensors, the drone can tell if the smell is coming from the left or right, allowing it to steer directly into the wind.

4. The "Vision" Upgrade: The Safety Net

Sometimes, the smell gets so faint or the wind gets so crazy that the drone gets confused. It might circle the source for minutes, trying to find the perfect gradient.

The Trick: The researchers added a camera as an optional "last resort."
How it works: The drone flies purely by smell 95% of the time. But once it gets very close to the source, it turns on its camera. If it sees the diffuser (the device releasing the smell), it says, "Aha! I found it!" and lands immediately.
The Analogy: It's like walking through a dark forest following a faint sound. You walk by ear until you get close enough to see the person, then you sprint the last few steps. This saved the drone time and battery.

5. The Big Test: The 200-Square-Meter Maze

The team tested their drone in a large indoor room (about the size of a tennis court).

The Setup: They placed a fan blowing ethanol (alcohol) from one corner. The drone started at the opposite end.
The Rules: No GPS. No pre-made maps. No human controlling the drone. It had to find the source on its own.
The Result: The drone succeeded 100% of the time. It found the source every single trial, usually in under two minutes.
- The "Metal Oxide" nose was faster.
- The "Electrochemical" nose was more precise but slower.
- Adding the camera made the drone slightly faster by cutting out the final "wandering" phase.

Why This Matters

This is a big deal because:

It's "Edge" Computing: The drone did all the thinking onboard. It didn't need to send data to a supercomputer in the cloud. This means it could work in places with no internet.
It's Open Source: The team gave away all their blueprints, code, and designs. Anyone can build this drone.
Real-World Applications: Imagine a drone that can find a gas leak in a building, locate a chemical spill in a disaster zone, or even help surgeons find infected tissue by smell during an operation.

The Bottom Line

The researchers proved that a robot doesn't need to be a genius with a massive map to find a smell. By giving it a simple, insect-like brain and a pair of "ears" (sensors) that listen to the wind, a small drone can chase ghosts and find exactly what it's looking for. They turned the impossible task of "smelling in the dark" into a reliable, repeatable science.

1. Problem Statement

Autonomous odor source localization (OSL) for aerial robots (UAVs) is a significant challenge due to the chaotic nature of gas dynamics. Key difficulties include:

Turbulent Airflow: Gas plumes are intermittent, fragmented, and highly sensitive to wind shifts, making them difficult to track.
Sensor Limitations: Olfactory sensors suffer from drift, hysteresis, long sampling times, and slow response rates compared to visual sensors.
Payload Constraints: UAVs have strict limits on weight and computational power, preventing the use of heavy sensor suites or complex external infrastructure.
Data Scarcity: There is a lack of standardized benchmarks and large-scale datasets for olfactory data, hindering the development of machine learning models.

Existing solutions often rely on pre-defined coverage patterns (gas distribution mapping), external positioning infrastructure (like UWB tags), or swarm coordination, which limits their applicability to single-agent, real-time, edge-computing scenarios.

2. Methodology

The authors propose a complete, open-source framework for a single UAV to navigate to an odor source using minimal sensing, trained in simulation and deployed on real hardware.

A. Hardware Design

Platform: A modified DJI Tello drone was selected for its open-source programmability and sufficient payload capacity.
Sensors:
- Olfaction: Two configurations were tested:
  1. Metal Oxide (MOX): Sensirion MICS 6814 sensors (fast response, 1-100 Hz, general sensitivity).
  2. Electrochemical (EC): Custom sensors using chronoamperometry with ionic liquids (slower but highly specific to ethanol).
- Placement: Sensors are mounted in a stereo configuration (analogous to insect antennae) on the front/top of the UAV to enable stereo sensing and avoid rotor wash interference.
- Support: Infrared (ToF) sensors for obstacle avoidance and altitude; a custom Olfactory Processing Unit (OPU) based on an ESP-32 microcontroller handles data acquisition and wireless telemetry.
Mechanical: Custom 3D-printed brackets and ballast (71g) were added to balance the center of gravity caused by the added electronics.

B. Simulation & Training

Digital Twin: A physics simulation (Simulink) models the control loop, while a Python/Gymnasium environment simulates the plume dynamics using Gaussian time-series processes and Dryden turbulence models.
Navigation Policies:
1. Olfactory Inertial Odometry (OIO): A rule-based algorithm inspired by insect behavior (surge/cast). It uses time-delay estimation between left/right sensors to determine plume angle and dual-timescale exponential moving averages (EMA) to detect plume "bouts" (onset/offset) while filtering noise.
2. Reinforcement Learning (RL): Two algorithms were trained: Q(λ) and Expected SARSA(λ). The latter was selected for deployment due to its robustness in handling "blank" air pockets (regions with no odor) and noise.
Vision Augmentation: A multimodal model called COLIP (Contrastive Olfaction-Language-Image Pre-training) was developed. It uses a CLIP-based vision encoder and a custom olfactory encoder to learn relationships between visual objects and chemical signatures. A lightweight Graph Associator (fully connected network) maps these embeddings to a shared latent space. Vision is used as a "last-mile" confirmation to stop the UAV once the source is visually identified, preventing wasted time chasing micro-gradients near the source.

C. Task Termination Criteria

Since the payload causes significant IMU bias, long-term position estimation is unreliable. The system determines it has found the source when:

It lands within 1 meter of the diffuser.
A statistical estimator calculates that the observed maximum concentration ( $m$ ) is within a 95% confidence interval of the true maximum (specifically, within 25% of the expected max).

3. Key Contributions

Open-Source System: A fully reproducible UAV system including custom hardware designs (PCB, 3D prints), firmware, simulation code, and datasets.
Minimalist Navigation: Demonstrates successful source localization without gas distribution mapping, external positioning (UWB), or prior environmental knowledge.
Stereo Olfaction-Vision: The first demonstration of stereo olfactory sensing combined with vision on a UAV, generalizing across two different chemical sensing mediums (MOX and EC).
Sim-to-Real Transfer: Successfully trained learning-based policies in simulation and deployed them on a real quadrotor in a large indoor environment (200 m²).
Edge-Optimized AI: Developed lightweight models (COLIP) that run entirely on the edge, avoiding the need for cloud-based LLMs or heavy inference.

4. Results

Experiments were conducted in a 200 m² indoor environment with an ethanol source. The UAV performed 5 trials for each of four configurations (MOX/EC sensors × OIO/RL algorithms, with and without vision).

Success Rate: 100% across all trials; the UAV consistently found the source.
Performance Metrics (Average Time):
- MOX + OIO: ~98.4 seconds (Best time: 71.6s).
- EC + OIO: ~112.2 seconds (EC sensors are slower due to sampling rates).
- Vision Augmentation: Reduced navigation time by ~4-14 seconds. Vision allowed the UAV to stop immediately upon visual confirmation, avoiding the "circling" behavior seen in olfaction-only runs where the UAV tries to maximize a flat gradient near the source.
Algorithm Comparison: The rule-based OIO showed more consistent timing than the RL approach. The RL model (Expected SARSA) was robust but slightly slower, likely due to over-optimizing paths based on eligibility traces.
Sensor Comparison: MOX sensors were faster, while EC sensors provided higher specificity but required longer sampling times.

5. Significance

This work bridges the gap between biological olfactory navigation (inspired by insects) and robotic autonomy. By proving that a UAV can localize a chemical source using only minimal, low-cost sensors and onboard computation, the authors establish a new baseline for edge-based olfactory robotics.

The framework is significant because it:

Removes reliance on expensive external infrastructure (UWB, GPS).
Provides a scalable, open-source foundation for future research in hazardous material detection, search and rescue, and environmental monitoring.
Demonstrates that "chasing ghosts" (navigating by scent) is feasible for machines, paving the way for multi-modal robots that can "smell" their way through complex, unstructured environments.