**4. Adaptive Bat Echolocation Behaviors Inspire Artificial Sonar Tracking Systems**

The echolocation and flight behavior of bats have been a source of inspiration for many technological advances, however, many key features of bat sonar tracking have yet to be realized in artificial systems. As described above, bats rapidly modify the spectro-temporal features of echolocation calls, and these adaptive sonar behaviors are fundamental to their performance in navigating complex environments while tracking and intercepting targets. Full implementation of these adaptive sonar behaviors, coupled with the use of wideband sonar signals, offers great potential for future technology applications. In this section, we present some examples in which bats have inspired technology thus far.

In 2010, a standard bat algorithm (BA) was proposed as a metaheuristic algorithm that uses similar processes of echolocating bats for global optimization [77]. The standard BA uses idealized behaviors of echolocating bats, which draws from a limited subset of parameters. These idealized behaviors or rules are as follows:


This algorithm iteratively updates the position and velocity of a virtual bat, using these three idealized rules. This updating allows for a more dynamic and efficient method for optimizing the processing of sensory information, allowing the BA to solve constrained and unconstrained optimization problems better than similar biologically inspired algorithms [78,79]. However, these idealized rules greatly simplify the components of adaptive echolocation, e.g., assuming that bats use a single sonar frequency, which changes with distance. The algorithm does not consider the bat's use of wideband FM signals or task-dependent behavioral measures at a given distance, such as preparing to intercept a target vs. flying by that target. The standard BA has been further developed to incorporate a directional bat algorithm (dBA), which improves performance in different types of environments and conditions, including premature convergence due to low exploration [80]. More recent advances in a binary bat algorithm (bBA) address traffic network determination problems [81] and parameter initialization to improve convergence velocity and accuracy [82]. Further integration of a more complete repertoire of adaptive behaviors of bats would continue to improve this optimization algorithm.

Robotic navigation has employed the Simultaneous Localization and Mapping (SLAM), which constructs and updates a spatial map of a novel, fixed environment, from both allocentric and egocentric perspectives, thus allowing an agent to navigate without *a priori* knowledge of the surroundings. In the last decade, there have been significant strides in creating reliable SLAM algorithms, however,

there are still limitations to these approaches. Sensor uncertainty, the processing demands of complex computations, and challenges of dynamic environments contribute to the current limitations of SLAM algorithms [83]. One approach to this problem is RatSLAM, which uses the computational models of the hippocampus in rodents to inform navigation in novel environments with ambiguous landmark information [84]. This biologically inspired approach to SLAM has yielded promising results, with increased place-recognition performance and recovery from path integration errors. Expanding on this biological model, Steckel and Peremans have proposed the use of the echolocating bat for a sonar-based model of SLAM [2], which can operate under conditions where optical information is reduced or unavailable. Previous SLAM systems with sonar capabilities required impractically large numbers of sonar measurement to converge on a functional map [2], but BatSLAM offers a new way to use sonar information more efficiently, to allow autonomous sonar-guided robots to navigate an environment. Like Yang's Bat Algorithm, BatSLAM draws inspiration from the bat's biosonar, to localize the positions of obstacles to generate an experience map and modify motor commands for path integration. Additionally, they use the physical structure of the bat's external ears to allow binaural sound localization, though they do not incorporate adaptive sonar signal design, head movements, or the ability of the bat to dynamically move each pinna independently, to amplify interaural differences (Figure 1). Combining directionality of sonar emissions and binaural echo reception of bats, Steckel and Peremans developed the Echolocation Related Transfer Function (ERTF) for spectro-spatial filtering, realizing a biomimetic sonar system that localizes multiple acoustic objects with wideband sonar [2]. It creates consistent maps of large environments that can converge over time, to relatively accurate metric maps, to support navigation. More recently, there have been advances to BatSLAM, which include odometric information and an acoustic flow model, which allows for a novel 3D sonar sensor [85], as well new optimization of 2D-experience mapping through the addition of an audio-aware perceptual hash with a closed-loop detection algorithm, using fixed CF and FM sonar signals [86]. These new enhancements to BatSLAM enable richer representations of complex environments, however, dynamic bat sonar adjustments offer many additional features that could be incorporated in future versions, for operation in more complex and dynamic environments.

Tracking algorithms have a myriad of uses, from surveillance [87] to biomedical applications [88]. While the accuracy of tracking algorithms has improved significantly in the last decade, they often fail to contend with background noise and clutter, which interferes with localization of a selected target. Kalman filters operate with iterative processes that aid in estimating the position and motion of a target and have been a standard for addressing the challenge of noisy target data. The addition of Kalman filters to tracking algorithms dramatically improves tracking fidelity and reduces interference by local maxima [89], particularly in linear systems. In nonlinear systems, extended Kalman filters also perform an iterative process with increased success [90], but concerns have been raised about inconsistent mapping and a penchant for underestimating covariance [91], particularly in handling sonar and vision data for the bearing of a target [92,93]. Improvements to complex tracking-condition algorithms have proven promising, such as application of multiple Kalman filters, which allows precise tracking of dynamically moving targets [94]. However, target tracking by artificial systems remains a challenge, and target interception success is still low.

Finally, biologically inspired approaches to sonar tracking algorithms have hailed some success, with iterative predictive algorithms approaching performance levels comparable to biological systems [95]. Both for biological and artificial systems, real-world tracking requires sensory input, which is then processed to output accurate pursuit of moving targets (Figure 4). Some models have implemented rudimentary behavioral features of bat flight trajectories and putative predictive tracking [74]. Behavioral research on bat sonar target tracking has provided valuable insights into the strategies these animals employ to perform complex tasks, including differential adaptation of their echolocation behavior with respect to moving targets and stationary obstacles [96], as well as Doppler shift compensation and discrimination [31], all while contending with different environmental constraints and conditions [27,30]. Empirical studies of adaptive and predictive sonar tracking

behaviors of bats [76], in conjunction with neurophysiological experiments, will provide insights to the computations employed by echolocating animals to carry out tasks under real-world conditions, and in turn provide further inspiration for new algorithms and neural networks that will improve artificial tracking systems.

**Figure 4.** Bats as a biological model to inspire tracking algorithms. Tracking moving targets requires sensory information that can be in the form of echoes (left/center panels) or visual stimuli (right panel). This information must then be processed by computations that allow for the prediction of future states (shown in gray). Man-made systems like Autonomous Underwater Vehicles (AUV) use sonar to track different moving targets, such as marine life and wildlife (left panel). Bats acquire discrete sensory information in the form of echo returns from adaptive sonar emissions; echo snapshots are integrated, to enable the prediction of the future position of a moving insect (center panel). These predictive tracking algorithms can be implemented in technologies that use sonar or other sensory modalities, such as drone videography of a bicycle race (right panel).
