Navigation: Where is that smell coming from?

Computational model reveals why pausing to sniff the air helps animals track a scent when they are far away from the source.
  1. Samuel Brudner
  2. Thierry Emonet  Is a corresponding author
  1. Department of Molecular, Cellular and Developmental Biology, and the Quantitative Biology Institute, Yale University, United States
  2. Department of Molecular, Cellular and Developmental Biology, the Quantitative Biology Institute, and the Department of Physics, Yale University, United States

Dogs, rodents and many other animals with a strong sense of smell often track a scent by keeping their nose to the ground, occasionally pausing to raise their heads and sniff something mysterious in the air (Jinn et al., 2020; Gire et al., 2016). However, exactly how alternating between these two behaviors helps animals navigate to the source of an odor remains unclear.

Airborne odors are transported by the wind, making them subject to the twisting and stretching of turbulent air motions. This results in animals downwind from an odor source being more likely to smell the odor intermittently, as air pockets containing the scent are interspersed with periods of clean air (Celani et al., 2014; Connor et al., 2018). Studies in insects suggest that animals surge upwind when they detect a smell in order to keep in contact with these turbulent odor plumes; when no odor is detected, they cast crosswind instead (Álvarez-Salvado et al., 2018; Demir et al., 2020; Kennedy, 1983; Flügge, 1934). Now, in eLife, Nicola Rigolli, Gautam Reddy, Agnese Seminara and Massimo Vergassola report how pausing to sniff the air when casting crosswind helps animals navigate towards the source of an odor (Rigolli et al., 2022).

To investigate how alternating behaviors impacts odor navigation, the team (who are based at institutes in France, Italy and the United States) designed a virtual search environment by simulating an odor dispersing downwind over a large area. The set-up created a challenging search scenario, including a low probability of encountering an odor pocket far from the source. Using machine learning, computer programs trained ‘artificial navigating agents’ to find the origin of the smell as quickly as possible (Sutton and Barto, 2019). During their search, these artificial agents were allowed to alternate between ‘walking’ while sniffing close to the ground and stopping to smell the air. Information gathered from these behaviors allowed agents to decide where to go next. After each attempt, the agents could use feedback about their previous search times to modify their strategy in the next trial. Although the researchers did not impose any explicit strategy or solution, agents reliably learnt that stopping to sniff the air sped up their search, even though it required them to pause.

Notably, trained agents mostly stopped to smell the air during the initial phase of their search before they had detected any odor. This suggests that alternating between ground and air sniffing helps agents to explore areas with dispersed levels of odor more efficiently (Figure 1A). Once agents successfully encounter an airborne cue, this signals that they have entered an odor rich zone and their rate of alternation drastically decreases.

The optimal strategy for finding the source of a smell.

(A) When tracking the source of an odor, animals alternate between walking while sniffing the ground (brown) and pausing to sniff the air (blue). Animals sniff the air more frequently when they are further away from the source and airborne cues are more dispersed (blue dashed line). As they get nearer and the density of the airborne cues increases (brown line), animals alternate less frequently and track the scent by sniffing close to the ground. (B) Rigolli et al. used a machine learning algorithm to identify the optimal strategy for tracking odors in the wind. They simulated an odor dispersing in the air (blue plume) and close to the ground (brown plume) and then trained artificial agents to find the source of the smell: the brown line indicates the trajectory agents took whilst sniffing the floor, and the blue circles represent where they paused to sniff the air. The algorithm revealed that the best way for agents to find the source of the odor was for them to alternate to sniffing the air when moving crosswind, and intersperse this with occasional surges forward until an odor was detected (blue star). (C) The simulation showed that agents displayed this alternating behavior less frequently as they moved closer to the odor source.

Image credit: Samuel Brudner; odor plumes in panel B are based on images by Nirag Kadakia and Mahmut Demir.

But how exactly does alternation speed up getting that first hint of a scent? Failing to detect an odor in a region indicates that the source of the smell is unlikely to be upwind of this area, eliminating the need to search there in the future. Odors disperse slower near the ground, and as a result do not reach as far as odors travelling in the air. Sniffing above their heads therefore allows agents to rule out larger upwind areas (if no smell is present), while also increasing the likelihood of detecting faint signals that are absent at ground level.

The simulation also revealed that during the early search phase, trained agents combined alternation with specific patterns of locomotion (Figure 1B). Agents moving crosswind sniffed the air more frequently than when they surged upwind. Rigolli et al. observed that this behavior helps the agents to rule out cross-sections of the simulated arena before moving upwind to gather evidence about a new region.

This cast-sniff-surge strategy involves many tradeoffs: casting over a wider distance takes longer but also eliminates a wider cross-section; sniffing in the air requires stopping and therefore losing time. Using a mathematical framework, Rigolli et al. show that optimally balancing these tradeoffs requires exploring the arena back and forth in an expanding crosswind zigzag, gradually casting across larger areas as the search progresses. Remarkably, these characteristics also appeared in trained agents which were not constrained to use a cast-sniff-surge approach.

Overall, Rigolli et al. demonstrate how, in theory, alternating between sniffing the ground and the air allows animals to efficiently search large areas for an odor source. Future studies should now test these predictions, for example examining if real animals do tend to alternate behaviors mostly in odor-poor regions. This work could also be applied to robotics, in particular to improve the exploratory behavior of drones used in difficult search-and-rescue operations.


  1. Book
    1. Sutton RS
    2. Barto AG
    Reinforcement Learning: An Introduction
    MIT Press.

Article and author information

Author details

  1. Samuel Brudner

    Samuel Brudner is in the Department of Molecular, Cellular and Developmental Biology, and the Quantitative Biology Institute, Yale University, New Haven, United States

    Contributed equally with
    Thierry Emonet
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-6043-9328
  2. Thierry Emonet

    Thierry Emonet is in the Department of Molecular, Cellular and Developmental Biology, the Quantitative Biology Institute, and Department of Physics, Yale University, New Haven, United States

    Contributed equally with
    Samuel Brudner
    For correspondence
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-6746-6564

Publication history

  1. Version of Record published: September 20, 2022 (version 1)


© 2022, Brudner and Emonet

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.


  • 187
    Page views
  • 31
  • 0

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Samuel Brudner
  2. Thierry Emonet
Navigation: Where is that smell coming from?
eLife 11:e82635.
  1. Further reading

Further reading

    1. Physics of Living Systems
    Ye Li, Shiqi Liu ... Yilin Wu
    Research Article

    Long-range material transport is essential to maintain the physiological functions of multicellular organisms such as animals and plants. By contrast, material transport in bacteria is often short-ranged and limited by diffusion. Here we report a unique form of actively regulated long-range directed material transport in structured bacterial communities. Using Pseudomonas aeruginosa colonies as a model system, we discover that a large-scale and temporally evolving open channel system spontaneously develops in the colony via shear-induced banding. Fluid flows in the open channels support high-speed (up to 450 µm/s) transport of cells and outer membrane vesicles over centimeters, and help to eradicate colonies of a competing species Staphylococcus aureus. The open channels are reminiscent of human-made canals for cargo transport, and the channel flows are driven by interfacial tension mediated by cell-secreted biosurfactants. The spatial-temporal dynamics of fluid flows in the open channels are qualitatively described by flow profile measurement and mathematical modeling. Our findings demonstrate that mechanochemical coupling between interfacial force and biosurfactant kinetics can coordinate large-scale material transport in primitive life forms, suggesting a new principle to engineer self-organized microbial communities.

    1. Immunology and Inflammation
    2. Physics of Living Systems
    Derek M Britain, Jason P Town, Orion David Weiner
    Research Advance

    T cells use kinetic proofreading to discriminate antigens by converting small changes in antigen binding lifetime into large differences in cell activation, but where in the signaling cascade this computation is performed is unknown. Previously, we developed a light-gated immune receptor to probe the role of ligand kinetics in T cell antigen signaling. We found significant kinetic proofreading at the level of the signaling lipid diacylglycerol (DAG) but lacked the ability to determine where the multiple signaling steps required for kinetic discrimination originate in the upstream signaling cascade (Tischer and Weiner, 2019). Here we uncover where kinetic proofreading is executed by adapting our optogenetic system for robust activation of early signaling events. We find the strength of kinetic proofreading progressively increases from Zap70 recruitment to LAT clustering to downstream DAG generation. Leveraging the ability of our system to rapidly disengage ligand binding, we also measure slower reset rates for downstream signaling events. These data suggest a distributed kinetic proofreading mechanism, with proofreading steps both at the receptor and at slower resetting downstream signaling complexes that could help balance antigen sensitivity and discrimination.