Machine Learning Approaches for Biosonar Systems: Investigating Dynamic Sensing in Complex Natural Environments

Abstract

Echolocating bats navigating dense vegetation extract sensory information from clutter echoes -- superpositions of reflections from numerous unresolved scatterers that create inherently stochastic signals. This dissertation investigates how both static morphological variations and dynamic shape changes in biomimetic receiver structures can encode reliable information within these random echoes, progressing from linear acoustic effects to nonlinear Doppler-based signatures. The first component of this research established that static pinna shape variations create consistent, discriminable effects on clutter echoes despite their stochastic nature. Using a biomimetic robotic pinna capable of producing ten distinct conformations, we recorded over 50,000 clutter echoes from artificial foliage agitated by fans. Deep neural networks successfully classified pinna shapes with 97.8% accuracy, demonstrating that linear beampattern variations survive projection onto random clutter backgrounds. This finding established the foundation for information encoding through morphological diversity. Building on these linear effects, the second component investigated whether dynamic pinna motions, which introduce time-variant and nonlinear Doppler signatures, similarly create reliable patterns in clutter. Using the same experimental apparatus actuated through different motion patterns at multiple speeds, we tested three signal types: constant-frequency (CF), frequency-modulated (FM), and combined CF-FM pulses. Classification accuracy varied dramatically with signal structure: CF signals achieved only 48.3% accuracy while CF-FM signals reached 97.2%, revealing fundamental differences in how linear and nonlinear effects interact with signal time-frequency characteristics. The progression from static to dynamic effects revealed a hierarchy of information encoding mechanisms: static shapes provide robust but fixed filtering, while dynamic motions enable adaptive, time-variant information encoding at the cost of increased complexity and energy. The superior performance with FM components suggests that broadband signals better capture motion-induced signatures, potentially explaining the evolution of diverse signal types across bat species. These findings have implications for both biological understanding and engineering applications. The demonstrated ability to extract reliable signatures from stochastic backgrounds suggests new approaches for acoustic sensing in complex environments, while the differential performance across signal types provides insights into the evolutionary pressures shaping biosonar systems. The work establishes machine learning as a powerful tool for uncovering subtle patterns in seemingly random biological signals. Methodologically, this research contributes to the growing intersection of deep learning and embodied artificial intelligence. The deep learning framework developed here treats neural networks as measurement instruments for probing information content in signals too complex for analytical characterization; a paradigm applicable beyond biosonar to any domain involving high-dimensional, stochastic data. More broadly, the findings validate morphological computation principles: the physical structure of sensors can actively encode information through geometry and dynamics, performing computations at the hardware interface that would otherwise require downstream processing. This embodied AI perspective where intelligence emerges from the interaction between physical structure, dynamics, and environment rather than residing purely in algorithms, suggests new design principles for autonomous systems operating in complex, unstructured environments where vision-based sensing fails.