Can this work underwater?

Yes, with suitable sensors and models for waterborne transport.

Does it always require a smooth concentration gradient?

No. Some methods are designed for intermittent detections and uncertainty.

Is infotaxis the same as moving toward the strongest smell?

No. It uses expected information about source location to guide search.

AI Glossary / Reinforcement learning and robotics / Intermediate

Odor source localization

Definition

Odor source localization is the task of identifying where an airborne or waterborne chemical plume originates, using measurements gathered by sensors and often mobile robots. It combines chemical sensing, movement, and inference about transport through the environment. Detecting a substance is only part of the task: a system must distinguish evidence of the plume from reliable evidence about the source's location.

What it means

A chemical sensor reports information where it is sampled, not a direct arrow pointing to the source. Wind, currents, turbulence, obstacles, and sensor response affect what is observed. A high reading may occur in a transported patch of material some distance from the release point.
In turbulent conditions, a plume can be intermittent rather than a smooth concentration hill. A robot may encounter repeated detections and gaps while moving toward the same source. Search methods therefore need to interpret both observations and the conditions under which they were obtained.

How it works

Choose sensing equipment suited to the target substance and operating conditions. Account for calibration, response time, cross-sensitivity, and the robot's ability to associate measurements with locations and times. Airflow or current measurements can provide useful context.
A search often involves finding the plume, following useful evidence, and deciding whether a source has been located. These activities may repeat when the robot loses contact with the plume. Some strategies use reactive movement rules; others maintain a probability distribution over possible source locations.
A probabilistic method updates its beliefs using a transport and observation model. Information-seeking approaches such as infotaxis select movement partly for the expected reduction in uncertainty, rather than simply seeking the highest immediate concentration. Their performance depends on how well the modeling assumptions fit the environment.
Define a source-declaration procedure and evaluate false declarations as well as search time and localization error. Use repeated trials with relevant airflow, obstacles, background signals, and sensor limitations. A successful controlled experiment does not establish dependable operation under every real-world plume condition.

Three practical examples

A research robot searches a controlled test area for a known chemical release and compares its estimated source position with ground truth.
An environmental monitoring system combines mobile concentration measurements with airflow information to investigate possible emission locations.
An underwater robot studies a dissolved chemical plume while accounting for current direction and changing transport conditions.

Think of it this way

Think of trying to locate a kitchen by catching occasional smells in a building. A strong smell in a corridor may have arrived through airflow, and losing the smell does not necessarily mean the kitchen vanished. You need to combine where and when you sensed it with evidence about how air moves.

Worked example

Consider three possible source locations, A, B, and C, with equal prior probabilities of one third. At one sampling location, an invented model assigns detection probabilities of 0.6 if the source is A, 0.3 if it is B, and 0.1 if it is C.
After a detection, multiply each prior by its detection likelihood: the unnormalized weights are 0.2, 0.1, and approximately 0.03333. Their sum is one third, so the posterior probabilities become 0.6, 0.3, and 0.1.
Suppose the robot then samples elsewhere and receives no detection. At that second location, assume detection probabilities of 0.8, 0.2, and 0.1 for A, B, and C. The corresponding no-detection probabilities are 0.2, 0.8, and 0.9. Multiplying by the current beliefs gives 0.12, 0.24, and 0.09. Dividing by their sum, 0.45, gives approximately 0.267, 0.533, and 0.200.
B is now the most likely candidate, but it is not confirmed. This calculation assumes the observation model appropriately accounts for sampling conditions and dependence. It illustrates Bayesian belief updating, not a full infotaxis policy or a realistic plume model. A non-detection can change beliefs without proving that a source is absent.

Illustrative case study

In this fictional case, Willow Monitoring tests a robot in a controlled ventilation chamber. The robot initially follows higher readings, but intermittent airflow makes its path inconsistent.
The team adds time-aligned airflow measurements and compares a probabilistic search method with the original strategy. It checks whether the observation model fits the chamber data and records runs in which the robot confidently selects the wrong candidate location.
The pilot report separates chemical detection, source localization, and source confirmation. The researchers treat results from the chamber as evidence for those test conditions and plan further evaluation before considering a different environment.

Common misconceptions and mistakes

The strongest reading is not necessarily at the source, and a missing reading does not necessarily mean no source exists. Turbulence and sensor response can both complicate interpretation.
Another mistake is assuming that any gradient-following rule works in every plume. Information-seeking methods also have assumptions and can fail when their model is unsuitable. The highest posterior candidate is not automatically a confirmed source.
Finally, odor classification and source localization solve different problems. Identifying the likely substance does not by itself determine where it was released.

Business use

Potential applications include environmental investigation, facility monitoring, and research inspection tasks. A practical evaluation should match the target substances, transport conditions, and required localization accuracy.
Measure false positives, missed detections, declaration errors, search duration, and maintenance requirements. Clarify how a candidate location is independently checked and how sensor limitations affect decisions. Robotic localization is one component of an operational investigation, not an automatic substitute for all other evidence.

Check your understanding

What happens after the first detection in the example?

Answer: A has probability 0.6, B 0.3, and C 0.1.

Why does the second non-detection favor B over A?

Answer: It is less surprising under B's assumed model.

Why is B not confirmed at the end?

Answer: Its posterior is only about 0.533, and the calculation depends on an assumed observation model.

Related terms

Bayesian inference
chemical plume
gas sensing
infotaxis
active sensing
source declaration
anemometry
mobile robot localization

Learn next

Learn next how chemical sensors respond over time and how turbulence shapes plumes. Study Bayesian source estimation and information gain, then compare methods under repeated, controlled tests with known source positions.

Information gain

Sources

Vergassola, Villermaux, and Shraiman, Infotaxis as a Strategy for Searching Without Gradients: https://www.nature.com/articles/nature05464 Original research, Effectiveness and Robustness of Robot Infotaxis for Searching in Dilute Conditions: https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2010.00001/full

Previous: Odometry Next: Off-line SPC