Fig. 3: Emergent plume tracking behaviour can be decomposed into distinct behaviour modules.

a,b, A successful (a) and an unsuccessful (b) plume tracking episode (RNN agent 3) showing three distinct behaviour modules: tracking (green), lost (red) and recovering (purple–blue). c, Kernel density estimates show data aggregated from an equal number of successful and unsuccessful constant-wind-direction plume tracking episodes (N time steps, E episodes). Plots reveal differences between the three behaviour modules across key behavioural measures: Head direction: head-direction densities are concentrated around ±180°, a signature of zig-zagging but mostly upwind movement. Angles are measured anticlockwise, with 0° indicating directly downwind. Density estimates for drift in the x direction (Δx) and y direction (Δy) per time step show how tracking is characterized by primarily upwind (negative x-direction) movement in both tracking and recover modules, but less so in the lost module. y-direction movements are notable in the tracking and recovering modules, corresponding to more complex turning behaviours, but minimal in the lost module. Turn action: left/right turning movements are balanced in the tracking module as the agent closely tracks the edge of the plume, but it is biased towards clockwise movements in the other two modules, especially the lost module. Move action: the agent substantially modulates its forward movement speed in the lost module only. Stray distance: the agent strays from the plume minimally in the tracking module, but substantially otherwise. Empirical distributions of course direction suggest that agents track the plume with respect to the plume centreline rather than the current wind direction. d–f, Kernel density estimates of an agent’s course direction relative to the local plume centreline (solid blue) and to the current wind direction (dashed orange) in three plume configurations. ±180° means antiparallel movement with respect to the plume centreline, or exactly upwind movement with respect to the wind direction. d, Constant wind configuration: the two course-direction distributions are equivalent. e,f, A substantial proportion of time is spent at an angle (≈45° angle for the switch-once configuration, e; arbitrary angle for the switch-many configuration, f) to the wind, but actually aligned (antiparallel) with the plume centreline. Bottom row panel titles indicate how many time steps and how many successful episodes were summarized in each plot.