Sector search strategies for odor trail tracking

Abstract

Terrestrial animals such as ants, mice and dogs often use surface-bound scent trails to establish navigation routes or to find food and mates, yet their tracking strategies are poorly understood. Tracking behavior features zig-zagging paths with animals often staying in close contact with the trail. Upon sustained loss of contact, animals execute a characteristic sequence of sweeping "casts" -- wide oscillations with increasing amplitude. Here, we provide a unified description of trail-tracking behavior by introducing an optimization framework where animals search in the angular sector defined by their estimate of the trail's heading and its uncertainty. In silico experiments using reinforcement learning based on this hypothesis recapitulate experimentally observed tracking patterns. We show that search geometry imposes limits on the tracking speed, and quantify its dependence on trail statistics and memory of past contacts. By formulating trail-tracking as a Bellman-type sequential optimization problem, we quantify the basic geometric elements of optimal sector search strategy, effectively explaining why and when casting is necessary. We propose a set of experiments to infer how tracking animals acquire, integrate and respond to past information on the tracked trail. More generally, we define navigational strategies relevant for animals and bio-mimetic robots, and formulate trail-tracking as a novel behavioral paradigm for learning, memory and planning.

Adaptive algorithms for shaping behavior

“…The agent perceives the exact magnitude of odor at its current location, which is represented in the magnitude of the green color channel in each image observation. This is in contrast to Reddy et al [71], which uses a Poissonbased odor detection model. See Figure 4a for a plot of an example trail, including the trajectory of a successful agent.…”

“…To construct naturalistic trail geometries, we use the procedure described in Reddy et al [71]. Trail characteristics are modulated with the following parameters:…”

“…Online: α = 0.05, β = 0.Naive: α = 0.16, β = 3.Window: α = 0.26, β = 4.83, k = 10 • Sampling: α = 0.1, k = 3 7. Trail trackingTrails are generated using generalized worm-like chain ensembles, using the procedure described in Reddy et al[71]. The parameters used to sample each trail are…”

Emergent behavior and neural dynamics in artificial agents tracking turbulent plumes

“…They trained this model using DRL to solve four tasks and then analyzed the virtual rodent's emergent behavior and neural activity, finding similarities at an abstract level between their agent and observations from rodent studies. Reddy et al [2021] studied the trail tracking strategies of terrestrial animals with one (e.g. one antenna) or two (e.g.…”

“…We build on the approach of these recent papers that study artificial agents solving neural inspired tasks, and our work is also distinct in several key ways. First, we simulate a more computationally challenging task than those tackled in Reddy et al [2021] and Rapp and Nawrot [2020], because our odor environment is configurable, dynamic, and stochastic. Second, we have made several simplifications and abstractions that make analysis more tractable, so that we may focus on the general principles behind plume tracking.…”