Arjun MJ via Unsplash (CC0)
Many animals use odors to locate mates, food, and to avoid danger. Unlike light, which travels in straight lines, odors are carried by turbulent air or water, leading to intermittent whiffs separated by long gaps with no detectable scent. These patchy odor landscapes can make it difficult for animals to decide which direction to move in.
Despite these challenges, animals are remarkably good at using odors to navigate. While previous studies have modelled this behavior computationally, the most principled models often relied on complex concepts of memory, that were not directly interpretable. In particular, what must be remembered about past odor detections and for how long remained unclear.
To investigate this, Rando et al. developed an algorithm that enables agents to learn to navigate by trial and error, responding only to a short excerpt of past odor detections. Agents had no prior knowledge about the odor nor access to spatial information, other than their ability to orient relative to the wind. The simulated environment mimicked realistic odor plumes in turbulent air and the algorithm was given a short-term memory to track changes in a limited set of specific odor-related signals over time.
Analysis showed that there is an optimal length of memory that helps the agent ignore temporary gaps in the odor signal while still recognizing when it has fully exited the plume. This allowed the agent to activate a strategy to return to the scent plume only when truly necessary. When it was allowed to learn behavior both within and outside the plume, it performed better than when using fixed strategies based on animal behavior. Interestingly, the learned strategy often resembled the casting behavior, seen in flying insects, which involves a side-to-side search in the crosswind direction to relocate odor plumes.
Overall, the work of Rando et al. shows that simple odor signals and a basic form of temporal memory are enough to learn effective navigation in turbulent environments with no prior knowledge of the odor environment. The algorithm performed reliably, reaching the odor source in 90% to 100% of trials. These findings help explain how animals might use short-memory of odor to navigate in space, even in unknown or variable environments and could be used to develop search algorithms for robots in complex real-world settings like disaster zones or polluted areas.
