Abstract
Abstract
Birds routinely solve high-speed, three-dimensional collision-avoidance problems that remain a central challenge for autonomous micro-aerial vehicles (MAVs). While pairwise avoidance is well studied in budgerigars, comparatively little is known about how avoidance kinematics change when more than two birds must simultaneously resolve a potential collision—the regime in which distributed control becomes attractive for MAV swarms. Using an openly available stereo-camera dataset (n = 150 1v1, n = 80 2v2, n = 30 3v3 trials), we extracted nine kinematic indices per encounter and tested for group-size effects using dual statistical frameworks supplemented by stratified, subsampling-robustness and noise-floor sensitivity analyses. The two most robust findings are (i) that the optic-flow time-to-contact margin τ is qualitatively conserved across group sizes (the time-course is preserved while the absolute margins compress with group size), and (ii) that the minimum inter-bird separation does not shrink with group size and in fact grows mildly. By contrast, evidence that birds in the 3v3 condition react later, modulate forward speed less, or fly more curved paths is conditionally supported at best: the reaction-distance effect replicates in only 19.5% of subsampled iterations, and the path-complexity effects are significant only after trial-level aggregation. The agility metrics (peak lateral acceleration and minimum turn radius) are heavily confounded with tracking noise; a majority of 3v3 encounters fall below the empirical noise floor and these values should not be used as an engineering target. These observations are consistent with—but do not strongly mandate—a bioinspired design hypothesis in which a fixed τ-based visual policy is paired with a group-size- adaptive reaction threshold. A proof-of-concept multi-agent simulation of this controller resolves pairwise head-on encounters reliably (100% collision-free) but degrades sharply with group size and density, indicating that a purely reactive single-threat policy is insufficient for swarm-scale operation and that an explicit multi-threat coordination mechanism is required.