Tracking Odor Plumes With AI Agents Using A Deep Reinforcement Learning Model

The extraordinary talents of animals have long served as a source of inspiration for scientists and engineers who have worked to reverse engineer or mimic these abilities in bots and artificial intelligence (AI) agents. One of these characteristics is the capability of some animals, especially insects, to focus on the source of particular scents of interest (such as food or mates), often over great distances. Flying insects frequently carry out this kind of monitoring across great distances in search of food or mates. It is strenuous to track an odor plume and find its source when the wind and plume data need to be corrected. Experimental research has been done on several elements of this unusual behavior and the brain circuitry underlying it. 

To better comprehend this astonishing capacity of flying insects, researchers from the Universities of Washington and Nevada, Reno, have developed a novel strategy employing artificial neural networks (ANNs). Their work, just published in Nature Machine Intelligence, is an example of how artificial intelligence generates novel scientific discoveries. In addition to being an essential biological skill, plume tracking is a remarkable example of natural intelligence because it requires the integration of memories about both current and past odors and the processing of sporadic or imprecise olfactory cues and wind sensory signals. This processing enables the insects to adapt their flight trajectories quickly.

Researchers may be able to develop more effective robots that can find and follow dangerous gas leakage, fires, and other environmental risks if olfactory plume tracking can be successfully replicated in robots or artificial agents. Artificial neural networks (ANNs) that have been trained on labeled data have become widely used by neuroscientists to research and simulate biological processes. In their research, simulations were utilized by the scientists to train ANNs rather than labeled data.

The researchers first created a model of an odor coming from a source inside a windy arena with a total size of about 120 m² to train their plume-tracking agents using DRL. Their agents were rewarded when they discovered the smell’s origin.  However, if they lost sight of the olfactory plume and fled the arena, they were “penalized.” The simulator’s ability to produce plumes with different odor concentrations and wind patterns gave the scientists an extra edge in observing how the agent would act in specific circumstances. The researchers’ findings indicate that their model might reproduce the molecular mechanisms that regulate animal olfactory plume tracking. The researchers produced plume configurations that might be imitated in upcoming practical wind tunnel tests. They established various theories about how artificial agents would act when following plumes in varying wind conditions thanks to these simulations. They paid particular attention to situations where the wind regularly changes directions.

ANN agents can be reverse-engineered to gain a greater understanding of how they function in addition to enabling significant technological breakthroughs, which may, in turn, influence neuroscience research. Thus, neuroscientists may also employ this model to investigate the molecular mechanisms underlying olfactory plume tracking.

The researchers envision their model influencing the development of autonomous agents that can track odors for future use in environmental monitoring, search-and-rescue missions, and other applications. By increasing the physical and biological integrity of their simulations and agents in their subsequent research, scientists hope to further enhance their model and make it more accurate at simulating actual olfactory plumes. Additionally, the scientists want to mimic other flying insects’ physiological traits and capacities.

Check out the Paper and Reference Article. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 14k+ ML SubReddit, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.