Bioinspiration from bats and new paradigms for autonomy in natural environments

Achieving autonomous operation in complex natural environment remains an unsolved challenge. Conventional engineering approaches to this problem have focused on collecting large amounts of sensory data that are used to create detailed digital models of the environment. However, this only postpones solving the challenge of identifying the relevant sensory information and linking it to action control to the domain of the digital world model. Furthermore, it imposes high demands in terms of computing power and introduces large processing latencies that hamper autonomous real-time performance. Certain species of bats that are able to navigate and hunt their prey in dense vegetation could be a biological model system for an alternative approach to addressing the fundamental issues associated with autonomy in complex natural environments. Bats navigating in dense vegetation rely on clutter echoes, i.e. signals that consist of unresolved contributions from many scatters. Yet, the animals are able to extract the relevant information from these input signals with brains that are often less than 1 g in mass. Pilot results indicate that information relevant to location identification and passageway finding can be directly obtained from clutter echoes, opening up the possibility that the bats’ skill can be replicated in man-made autonomous systems.


Introduction 1.Levels of challenges for autonomous systems
Central to most definitions of autonomous systems is the ability to accomplish tasks without direct human intervention [2,12].In order to achieve the desired objectives reliably in a variable physical world, an autonomous system needs to adapt to changes in its environment [46,56].Sensing capabilities are key for this ability to adapt since they provide the information that is needed direct the adaptations.As a result, sensing has been included explicitly in several definitions of autonomy [23,33].
So far, successful applications of autonomous systems have been limited to environments that can be controlled almost completely to make autonomous operation easier in general and more reliable in particular (figure 1).Examples are manufacturing lines [24] and warehouse operations [28], where almost all aspects of the environment can be controlled and set up to facilitate the operation of autonomous systems.This could include sufficiently precise limits on the position, orientation, and characteristics of objects to be manipulated by a robot as well as control over environmental conditions such as lighting.Furthermore, the workspaces can be structured with readily recognizable and unequivocal markings for fiducial points and borders.Similarly, the presence of interfering objects or signals could be kept to a minimum.Finally, the autonomous systems can be provided with large amounts of a priori information on the predetermined structure of the environment.Due to all these favorable factors, such 'fully controlled' conditions could hence be seen as the least demanding level of autonomy.
In environments where such far-reaching control measures cannot be implemented, the general performance and in particular the reliability of autonomous systems quickly deteriorates (figure 1).An example for this is autonomous driving [68], where the environment could be characterized as 'partially controlled': road traffic is controlled by a fixed set of rules as well as uniform standards for aspects such as road markings and signage.However, there are factors such as weather phenomena and the behavior of other traffic participants that are not entirely predictable and hence continue to pose a challenge to the autonomous capabilities of selfdriving cars [39].'Partially controlled' environments could hence be seen as falling right on the edge of current autonomous capabilities.
The final and most demanding level of autonomy can be found in complex natural environments (figure 1).While these environments are still constrained by the fundamental laws of nature, they have to be taken 'as is' and hence cannot be controlled for the purpose of making autonomous operation easier.Between this extreme case and the partially controlled environments that were just described lies a continuum of 'semi-natural' environments where some control is being exerted by humans, e.g. in agriculture and forestry.Dealing with the numerous unpredictable aspects of naturally or even 'semi-natural' environments successfully calls for autonomous systems with powerful mechanisms for adaptation to uncontrollable environmental factors.To realize these adaptation mechanisms, the importance of the sensing component of an autonomous systems is likely much more critical than in the fully controlled and even in the partially controlled situations.This is because any environmental condition that is relevant to autonomous operation but not known in advance has to be determined from sensory data.To exacerbate the situation, it is to be expected that the greater the complexity of an environments, the greater the complexity of the sensory data that is obtained from it.Hence, it can be expected that the importance of the sensing component of an autonomous system increases from fully controlled over partially controlled and finally to completely uncontrolled environments.As a consequence, autonomy in uncontrolled, complex natural environments is not only likely to pose the greatest challenges for autonomy in general but also for the supporting sensory systems in particular.

Importance of autonomy in natural environments
While reliable autonomy in complex natural environments remains well beyond current technical capabilities, achieving this goal in the future could have farreaching impacts on critical societal issues: many of humanity's most fundamental needs have to be met in outdoor settings of varying complexity.This includes the production of food and regrowable resources as well as maintaining healthy ecosystems and (national) security.A substantial number of the tasks in these areas could benefit at least as much from automation as has been the case for the traditional manufacturing and logistics areas that have already been revolutionized by automation.
An example for a field that could benefit greatly from outdoor automation techniques is agriculture.Here, autonomous systems could not only help to alleviate labor shortages, but could also advance a precision approach to agriculture where water, fertilizer, and herbicides/pesticides are dispensed in a highly adaptive fashion that is based on precise knowledge of local conditions [27].This approach could hence not only lead to greater productivity but-at the same time-would result in more efficient usage of resources and a better stewardship of the environment.Similar scenarios can be found in forestry, e.g. for autonomous monitoring of forest health and growth [11], detecting forest fires [36], as well as for seeding new trees in forests that lack accessibility and the ability to regenerate naturally [55].
Beyond agriculture and forestry, autonomous systems could play a key role in the conservation of biodiversity and healthy ecosystems [31].Taking sound decisions on environmental conservation requires detailed data, e.g. from surveys of the species composition in a given area [42].Devices for recording images ('camera traps' , [48]) and the sound [30] from animal vocalizations have already made a very noticeable impact on how much information on the species composition in a natural environment can be gathered [57].However, while these devices have some minimal amount of autonomy in that their recordings can be triggered by, e.g.motion or sound, they remain immobile and have very little ability to adapt to their task and their surroundings.For example, if a conventional camera trap has been positioned in the wrong place or oriented in the wrong direction, it cannot recognize these mistakes and select a different position or orientation autonomously.A genuinely autonomous recording system would be able to adaptively search for target biological species in different places and continuously improve its search strategy to track down even very rare species.Improvements in the quality of data that autonomous systems are able to could go beyond recording the presence of certain species and lead to the detection of much more subtle reactions changes in the ecological context such as climate change or the spread of disease in animal populations.
Besides agriculture and ecosystems monitoring, environmental clean-up constitutes a pressing environmental issue [43] that would require a high degree of efficiency that can likely only be realized with some degree of automation.An example for this would be the removal of plastic waste from land, water ways of all sizes, as well the oceans.To address this issue, autonomous systems could be deployed to characterize the distribution of waste in natural environments, to discover the sources and the most heavily affected areas, and finally carry out thorough garbage collection on a large scale.The autonomous sensing capabilities needed to accomplish this would include navigation in complex natural environments, e.g.along heavily vegetated shorelines, while detecting plastic objects of highly variable sizes and shapes that could also be partially overgrown or buried.
Finally, many critical security tasks need to be carried out in complex and largely uncontrollable natural environments.Examples include protecting natural environments from threatening activities such as wildlife poaching and illegal logging or securing national borders in natural areas.Battle fields can also be regarded as a prime examples for complex and poorly controllable environments.Besides natural uncertainties due to factors such as terrain, vegetation, and weather, the uncertainties associated with operation on a battle field can be exacerbated by enemy actions.For example, global positioning system (GPS) can be seen as a tool to bring an environment at least partially under control by taking out the uncertainty that is related location.An enemy could then try to curtail autonomous operation on the battle field by interfering with the GPS signal through jamming or spoofing.Hence, it is critical for autonomous operation on a battle field to provide local sensory capabilities that are not susceptible to enemy interference.
A common thread that runs through all these application areas is that autonomous systems could have a transformative impact through their ability to collect sensory data that would greatly exceed what manually-controlled methods could achieve in quantity and scope as well as through the ability to take direct action autonomously.This highlights once again the importance of sensing in these autonomous applications.

Conventional approaches
A common theme that runs through the conventional approaches to sensing for autonomous operation in complex natural environments has been to counter the lack of controllability and predictability with the collection of large amounts of data.Examples of this approach are the use of vision [49], including monocular [49], stereo [50], and depth cameras (red green blue-depth (RGB-D) cameras, [72]), as well as radar [71] and laser scanning (LIDAR, [38]).The common goal behind applying any of these sensing methods is usually to capture sensory data that would-at least in principle-suffice for building a digital model of the environment that is detailed enough to support planing of autonomous actions.A LIDAR scan, for example, could provide a digital three-dimensional model of all boundary surfaces in an environment that could then be used to perform path planning in a virtual equivalent of the real environment.Furthermore, the detail provided on the scanned objects, e.g.their shapes and surface textures, could be used to recognize objects for the purpose of identifying targets of interest.
Due to the large volumes of data involved, these approaches could be characterized as 'big data' strategies designed to ensure that no potentially important bit of information is missed.However, it should be pointed out that in this way dealing with the critical problem of extracting the relevant information is only postponed from the sensory data acquisition to the (digital) data processing stage.Hence, while sensory systems such as LIDAR that indiscriminately 'digitize' an environment may offer something akin of a 'guarantee' that everything there is to be known is captured in the scan data (at least down to a certain resolution and subject to other limitations such as occlusions), there is no guarantee that this information can be extracted, interpreted, and put to use successfully-let alone efficiently.
A key disadvantage of postponing the extraction of useful information from the sensing to processing stage is that it requires moving large amounts of data through the system.To accomplish this, sufficiently large capacities for data transfer, storage, and processing have to be provided.This necessitates greater hardware complexity, which typically causes larger system sizes and mass as well as higher cost, and higher levels of power consumption.In addition, handling larger amounts of data causes longer latencies.Device cost, complexity, size, power consumption, and latency are all critical factors that can stand in the way of realizing autonomous systems that can perform useful tasks in natural environments: high unit costs prohibit use of large numbers of systems that would be needed to perform useful outdoor tasks at scale.A high power consumption limits the time a system can operate or necessitates a bulky energy source to meet these needs, e.g. in the form of a large battery or a solar panel.Similarly, large latencies will slow the system down and hence limit its ability to react to fast changes in its environment.

Bat-inspired approach
Biological sensing in general appears to follow a very different approach from the 'big data-collection' strategy that is common in man-made systems that are engineered to deal with complex and unpredictable environments.Biological systems often make do with much lower data rates than what a high-performance LIDAR sensor would produce for example.To accomplish this, they start extracting the useful information from the sensory data immediately at the acquisition stage.Hence, the data that is transmitted from the sensors to the brain is already highly processed.Images recorded by the retina, for example, undergo extensive spatial bandpass filtering and other transformations before being transmitted to the brain via the optic nerve [64].In engineering, this 'edge computing' approach [61] has also begun to take hold, but in biology it is pervasive.However, biological systems often go even further than current edge-computing approaches, where incoming sensory signals are processed shortly after having been transferred into the digital domain.The mammalian inner ear that is an important part of the bat biosonar system is a good example of this since the decomposition of the acoustic signals into a 'bank of bandpass filters' representation is carried out in the physical domain by virtue of the mechanical properties of the sensor (i.e. the basilar membrane in the cochlea [40]).Hence, the primary sensory cells (inner hair cells) have already the desired bandpass characteristics without the need for any neural computations.In addition to the edge computing, the neural signal processing that occurs in biology is also geared toward representing salient features as opposed to conserving the raw sensory inputs [32].
As will be outlined in detail below, bat biosonar could be seen as an extreme case of the biological approach to sensing in terms of the importance of edge computing and abandoning an indiscriminately detailed representation of the environment in favor of a focus on extraction and representation of a small set of salient features.Beyond exemplifying the biological approach to dealing with sensory data, bats also make an excellent model system for bioinspired autonomy due to the outstanding performance that the animals have achieved in operating complex natural environments: bats have highly active, often predatory lifestyles that play out in three dimensions andin many cases-amid the bewildering complexity of dense vegetation.Combined with the proven ability of many species to make fine discriminations about objects in their environments, bats could serve as a rich source of insights into potential sensing strategies that could transform many of the application area for outdoor autonomy given above.
While all bat species can use vision to at least a limited extent [19], bat species with sophisticated biosonar systems have the demonstrated ability to accomplish demanding tasks such as navigation and hunting based on biosonar as the sole source of sensory information on their environments [15].Early attempts to explain the sensory abilities of bats have drawn on sonar theory [6,7] for the estimation of fundamental parameters from traditional sonar operation such as target distance [53], direction [5], velocity [8], and identity [4,9,54].In these cases, the assumption is usually made that the desired information has to be extracted from an echo that has originated exclusively from the target of interest and has been corrupted by additive white Gaussian noise along the transmission channel [59] but is separated in space from any interfering signals (figure 2(a)).
The core assumption of a single-target origin for each of the received echoes facilitates an analytic treatment of many aspects of sonar estimation and is well suited for open-ocean underwater sensing where sonar has been commonly used.However, the singletarget assumption also makes this estimation framework a poor match for situations where a bat is navigating in dense vegetation and hence surrounded by sonar targets that contribute to the received echoes (figure 2(b)).Natural foliages typically contain many reflecting facets, e.g.leaves, that can each contribute to the echoes that a bat biosonar system receives when operating in a densely vegetated environment [45].In sonar operations, echoes that contain contributions from unresolved scatterers are classified as 'clutter' (figure 3) that is notoriously difficult to interpret [22].A potential remedy that may appear as obvious would be to first resolve the echo contributions of the individual scatterers and then perform the desired sensing task, e.g.localization or identification, on each of these resolved components.However, resolving the  scatterers in a foliage would be difficult to accomplish for any sonar system given the typically large numbers of the scatterers and their close spacing in range.
This is exacerbated by the inherent physical limitations on the resolution of a bats' biosonar systems: all bat species rely on ultrasonic wavelengths in the millimeter to centimeter range-presumably since shorter wavelength would result in too much atmospheric absorption in the transmission channel.At the same time, the sizes of the bats' ears are limited by what a small flying mammal can accommodate.As a result, the aperture diameters of bat ears are more than ten times the acoustic wavelengths that the animals receive [58].These unfavorable ratios between apertures and the respective wavelengths limit the ability of bat biosonar to form the narrow beams that would be needed to resolve targets that are closely spaced in direction, e.g. the leaves in foliage along the angu-lar dimension.As a consequence, bats navigating in or near dense vegetation will almost certainly have to treat the echoes they receive as clutter due to unresolved scatters.
Bats that have mastered navigation in cluttered environments (figure 2(b))-despite the physical limitations on their biosonar systems-may be a biological model for an alternative approach to the detailed world models that dominate the state of the art in autonomous systems.Since bats are prohibited from creating detailed world models of complex environments, it may be hypothesized that the evolution of their biosonar systems has been geared toward ways to obtain the relevant information directly from clutter signals, i.e. without a detour over creating a model that does serves the informational needs of the bats only indirectly.If this assumption is correct, it would make bats a valuable biological model for an alternative and much more parsimonious approach  [20]).The indicated bandwidth of 20 kHz determines the data rate needed for encoding the signal, whereas the center (peak) frequency of about 84 kHz does not affect the information rate.Adapted from [20].CC BY 3.0.
to the current paradigms for dealing with complex environments.
Beyond dealing with the physical limits on angular resolution, bats also appear to have mastered navigation in complex environments based on much lower sensory data rates than what is common in engineered systems that rely on sensory modalities such as stereo vision or lidar scanning (figure 4).Greater horseshoe bats (Rhinolophus ferrumequinum), a species known for its capability to hunt in dense vegetation, for example, have been reported to use biosonar pulses with slightly less than 20 kHz bandwidth (for the second, strongest harmonic).The bandwidth is here defined as the difference between the highest and lowest frequency in the signal in the spectrum of the strongest harmonic.It is hence independent of the carrier frequency of the pulses, e.g.slightly above 80 kHz in greater horseshoe bats.To determine the data rate that is needed to encode the signal, only the bandwidth matters and the carrier frequency does not need to be considered [52].Here, the Nyquist frequency corresponding to the bandwidth is used to estimated the data rate needed to represent a pulse.Furthermore, greater horseshoe bats have been reported to employ a duty cycle of about 66% when hunting in the wild [35], i.e. 66% of the time are filled with pulses and the remaining 34% are allocated to pauses within the pulses.The duty cycle lowers the overall data rate, since no data needs to be transmitted during the pauses.Finally, assuming an amplitude resolution of 8 bit (i.e. a signal-tonoise ratio of 48 dB which would be very good for faint echoes), this would correspond to a data rate of about 26 kB s −1 for one ear or 52 kB s −1 for the binaural system.By comparison, state-of-the-art LIDAR scanners can produce data rates of about 160 MB s −1 (e.g. three million points per second [34] and 57 bytes per point, using the LAS point data record format 4 including wave packet descriptions [10]).This means that the data rate generated by such a LIDAR sensor is slightly more than 3000 times that associate with the biosonar system of a greater horseshoe [10] bat on the hunt.These vast differences in the sensory data rates suggest that the evolution of bat biosonar has created extremely efficient mechanisms for representing all essential information on an environment in a low-rate data stream and then extracting it at the neural processing stage.Considering the low-data rates together with the severe physical limitation on angular resolution, it is highly unlikely that these encoding and extraction processes involve an indiscriminately detailed reconstruction of the environment as would be the case in navigation based on LIDAR scans or stereo images.Instead, it could be hypothesized that the critical information is extracted directly from the clutter echoes without any attempt to resolve the different scatterers that contributed to them and creating a model of their spatial arrangement.
Similarly, the small brain masses found in bats (typically less than 1 g [18]) suggest that the neural extraction of sensory information may be just as parsimonious as its encoding in the periphery of the biosonar system.Given general estimates for the neuron density of mammalian cortex tissue [29], the bat brain masses would correspond to a few ten million neurons.In contrast to this, the number of computational elements in man-made computing systems has been increasing for several decades (Moore's law, [41]) and has reached numbers that now vastly exceed the numbers of neurons in any bat brain.Wafer-scale computing engines, for example, have reached 2.6 trillion transistors [3].However, these numbers are not readily comparable, since each neuron in a brain could function more like a complete 'analog computer'-often with a large number of inputs [14] that could perform pattern recognition at the cellular level [37].In contract to this, the computational elements in a digital computer just function as binary switches, but are typically operated at very high switching frequencies that exceed the spiking frequencies of neurons by many orders of magnitude.Taken together, it appears that the superior performance of bats when it comes to autonomy in natural environment is realized with a computational effort that is a lot less than what engineered systems would typically muster for such a task.Hence, it may be hypothesized that the superior performance of bat biosonar when dealing with complex natural environments does not stem from superior computational power and on the contrary it should be possible to outperform the state of the art with much less computational resources.The apparent computational parsimony of bats (and probably many other neural systems across the animal kingdom), could hence offer an alternative to the current development trends in artificial intelligence (AI) systems that are often aimed at achieving better performance through increasing the number of computational elements and network parameters [1].
While the dimensionality and data rates of the sensory inputs of bat biosonar as well as the computational capacity of the bats' brain are much less than what can be found in man-made systems designed for similar purposes, bat biosonar has additional complexity in places where it is not found in engineered systems: technical systems tend to keep their (sensory) periphery simple, e.g.linear and time-invariant, in order to facilitate understanding and reversing its impact on the signals that pass through it.In contrast to this, the periphery of biosonar systems in bats can be much more complicated, i.e. strongly time-variant and nonlinear.Certain bat species with sophisticated biosonar systems that allow the animals to navigate and hunt their prey in dense vegetation, in particular, horseshoe bats (family Rhinolophidae) and Old-World leaf-nosed bats (family Hipposideridae), have highly dynamic baffle on the emission and reception interfaces.These bats emit their biosonar pulses through nostrils that are surrounded by megaphonelike baffles that are actuated by a musculature that consists of 14 muscles [60].The noseleaf actuations can take place while the biosonar pulse is being emitted [21] and result in changes to the acoustic characteristics of the pinna.Similarly, the outer ears of these species are actuated by about 20 muscles [51] that can create a variety of motions which span rigid rotations as well as deformations [26,67].Linear characterizations of these motions have shown that changes to noseleaf and pinna shapes result in the encoding of additional sensory information, in particular with respect to target direction [44].In addition, the pinna motions in some bats have been shown to be fast enough to impart Doppler shifts on the incoming ultrasonic waves as they are diffracted into the ear canal by the moving pinna surface [66].
Together with the already hard to interpret nature of the clutter echoes, the complexity in the periphery of these bat biosonar systems is only likely to add to the difficulty that engineers would encounter when trying to make sense of the ultrasonic signals that enter the ear canal of a bat from the respective species in a complex natural environment.However, if the conventional reconstruction paradigms are abandoned in favor of direct estimation of relevant parameters, the biosonar echoes could provide an opportunity for deep learning techniques that have the demonstrated ability to discover complex patterns in data that were previously not accessible [62,63,66,69,70].

Pilot results
A combination of a biomimetic periphery inspired by bats (figure 5) with deep-learning echo analysis has shown promising pilot results that are relevant to accomplishing sensory tasks that could be key to autonomous navigation in natural environments.An example is the ability of finding passageways in natural vegetation [62,63].Detecting narrow gaps in foliage poses a challenge to bat biosonar, because the animals have wide beams [58].When looking at a foliage with a wide sonar beam, the echoes originate from a likewise large sonar footprint on that foliage.The received echo is the superposition of reflections from all scatters (e.g.leaves) with this sonar footprint.Hence, the presence of a gap that is small compared to the overall size of the sonar footprint will result only in the removal of small portion of the scatters and consequently will manifest itself in a small reduction of echo amplitude.This will make detection of narrow gaps with a wide-beam sonar using echo amplitude too unreliable to support navigation in dense vegetation.This problem is as relevant to engineering as it is to bats: the width of the biosonar beams seen in bats are limited by the ratio of the size of the biosonar apertures (e.g.mouth or noseleaf for emission and pinna for reception) and the wavelengths employed by the animals.If an engineered system was to form beams that are substantially narrower than those seen in bats, it would either need apertures that are much larger than those of the bats or employ wavelengths that are much shorter than those used by the bats.Substantially larger apertures (e.g. by an order of magnitude) would be a poor match for small drones, because of their size, weight, and the drag they would produce.Shifting to substantially shorter wavelengths may not be an acceptable solution either since it would result in much stronger attenuation of the signals due to atmospheric absorption of sound being proportional to the square of frequency.The large attenuation of the signals with distance would in turn result in much shorter sonar operation distances that may no longer meet the requirements for basic navigation or other tasks the system needs to accomplish.
Wide (bio)sonar beams are an impediment to the reliable detection of narrow gaps in foliage, because the energies of the echoes that are received when a gap is present and when it is absent can be expected to be poorly differentiated, i.e. the distributions of the respective amplitude values should overlap widely.This is because the mean values will be similar for gap and no-gap conditions.At the same time, the foliage-echo energies should also show a substantial amount of variability due to the random nature of the clutter echoes.Experiments with classifying biomimetic foliage echoes as originating from closed foliage or foliage with a gap have confirmed that while a performance well above chance level can be achieved based on echo energy, the performance is not nearly good enough to support reliable navigation.For example, in such experiments a false alarm rate (i.e.closed foliage classified as a gap) of about 10% was combined with a hit-rate (i.e.gap correctly detected) of about 50% [63].This would mean that a bat or a drone relying on echo amplitudes for gap detection would crash into closed foliage in one out of ten approaches to foliage and would be able to find only half of the existing gaps.
A deep-learning classifier that was trained to detect gaps in foliage based on echo spectrograms using a network architecture (convolutional neural networks) that has been inspired by the way the early stages of the brain's visual system process images.These networks were able to perform the gap-detection task much better than an energybased detector resulting in performance levels thatwhile not perfect-would much more commensurate with the demands of reliable navigation [62,63].Furthermore, it was possible to use a transparent AI method on the deep neural network to identify which parts of the foliage-echo spectrograms its classification decisions were based on.The method, classactivation mapping [63], identified the rising flank of the echoes as the region that contained the inform-ation on the presence of a gap.With this insight, it was possible to carry out the classification based on an ad-hoc neuromorphic spike-timing code, where the shape of the echoes' rising flank was represented by the times when the echo amplitude crossed a given threshold for the first time.This result is hence significant in two ways: first, it demonstrates that the clutter echoes contain information that is immediately relevant to navigation in natural environments.Second, it also shows that this information can be extracted in a highly parsimonious fashion.
Besides the ability to find passageways, the abilities to identify a location and build a map of the environment are critical for autonomous navigation.An analysis of pilot data has indicated that the identification of locations based on biomimetic echoes is possible on at least two different size scales [69,70]: it has been demonstrated that single biosonar echoes contained enough location information to distinguish ten different vegetated sites that were distributed over a region with approximately 40 km diameter [69].Furthermore, it was possible to classify two different tracks along with the data was collected at each of the sites.The later indicates that the biomimetic echoes contain information that is suitable to identify locations on a much finer scale.This was further corroborated by second study that showed that small numbers of biomimetic sonar echoes could support classifying small patches of vegetation that corresponded to a localization accuracy of about 6 m [70].However, since this experiment was limited by the accuracy of the GPS receiver that was used to provide the location reference, the accuracy that the biomimetic echoes could provide might have been even better than this value.
Nonlinear effects of the unique dynamics in the pinna of horseshoe and Old-World species could also play a key role in encoding spatial information that is suitable to support autonomy in complex natural environments.A pilot study has found that Doppler shift patterns created by a biomimetic have been found to be direction-depended and able to support determining the direction of a sound source with high accuracy and based only on a single pressure receiver and a pulse containing only a single carrier frequency [66].By contrast, conventional pressure-based localization of a sound source requires either multiple receivers or a broadband source signature.However, while this new source-localization paradigm certainly goes beyond the state of the art, it is not part of the sensory skill set that would be needed specifically to deal with complex natural environments.Hence, it remains to be seen if these Dopplermediated nonlinear effects could also play a role for clutter echoes generated in natural environments.

Future outlook
The study of biosonar and biomimetic sonar inspired by bats has produced a set of interesting sensory capabilities, some of which (passageway finding and location identification) could be directly applicable to the problem of autonomous navigation.Others, such as Doppler-shift patterns induced by pinna motions, go beyond the state of the art in acoustic sensing, but it is not yet clear whether they could play a role in encoding sensory information that is useful for complex natural environments.
Since all results so far have been obtained in relatively small pilot studies, the generality of these findings also needs further investigation.For example, it could be that some natural environments are substantially more homogeneous than others, which could effect the resolution that can be achieved when navigating based on natural landmarks.
Besides establishing that navigation in natural environments based on biomimetic sonar is feasible in principle, the question of how efficiently these skills could be learned when mimicking bats has not received any attention so far.Bats are known for fast development in general, which includes pups being born with up to 43% of the adult body mass [47].Juvenile greater horseshoe bats (R. ferrumequinum) have been found to undertake their first foraging flights at 24-28 days of age [17].Combining these short maturation periods with the comparatively low data rates, e.g.around ten pulses per second for greater horseshoe bats [35], associated with biosonar and the limited activity duration each night, e.g. less than 4 h after emergence [25], it could be estimated that a hypothetical juvenile bat that fits these example specification would have received a total of less than four million echoes, before it needs to function independently.However, it should be noted that more detailed investigations of the biosonar behavior of juvenile bats, especially with respect to activity times and pulse rates across these periods would be needed to arrive at a reliable estimate.Regardless of the exact number of echoes that a juvenile bats has at its disposal for learning, this echo data set would have to cover the entirety of the sensory tasks that the bat needs to perform at this stage, e.g.controlling various flight maneuvers as well as navigation and prey capture in different environments.
Assuming an average pulse duration of 60 ms [65] and a monaural data rate estimate of 40 kB s −1 (i.e.Nyquist sampling rate for 20 kHz bandwidth), four million echoes would correspond to 19.2 GB of binaural uncompressed data.From these ballpark estimates, it is to be expected that bats have to learn each of all its biosonar-based tasks based on far smaller amounts of data than is used in the state of the art for training deep neural networks.The training data set for the language model GPT-3, for example, was extracted from 45 TB of compressed plain text [13].Hence, even if the estimate for the number of echoes that has been used here is wrong by an order of magnitude, there would still be a vast gap between the amounts of training data that each system requires.
Besides the size of the training data set, it would also be worthwhile investigating whether bats have effective strategies to learn their environments that go beyond the supervised learning based on randomly acquired echo samples that has been used in the pilot studied so far.For example, it may be hypothesized that bats have smart and efficient strategies to sample a new environment sequentially which could insure that the received echo samples are particularly informative.
To facilitate system integration, future research efforts need to be directed at how to integrate the specific nature of the biomimetic clutter echoes into navigation paradigms.This could be realized through the adaption of existing frameworks, e.g.simultaneous localization and mapping [16] or through the creation of new paradigms that are custom designed to match the unique signal properties of clutter echoes and their spatial distribution in complex natural environments.
Finally, future work should investigate the challenges and opportunities associated with using biomimetic ultrasound on mobile platforms that are meant to operate in complex natural environments such as drones, terrestrial, or underwater vehicles.In-air sonar has a only a limited range which is not suitable for vehicles that have to move fast in large open spaces.In such situations, sonar could be replaced with radar, but similar bat-inspired approaches to accomplishing autonomy could be used despite the change in the physical modality used.In underwater application, sonar has a much longer range than electromagnetic waves, hence, this would not be an issue.For in-air system that are meant to operate in dense environments, where the nearest targets are never far away, the short range does also not impose any limitation.

Figure 1 .
Figure 1.Levels of difficulty for the operation of autonomous systems.

Figure 2 .
Figure 2. Traditional sonar operation versus the sonar environment of a bat hunting in dense vegetation: (a) sonar target with range, direction, and identity against a backdrop of clutter targets such as vegetation (light gray); (b) bat navigating in dense vegetation where all navigation information has to come from clutter-producing targets (light gray).

Figure 3 .
Figure 3. Examples of clutter echoes from natural vegetation (foliage) triggered by a biomimetic sonar pulse that consists of frequency-modulated and constant-frequency (narrowband) portions as can be found in horseshoe bats.

Figure 4 .
Figure 4. Example biosonar pulses of a horseshoe bat (Rhinolophus capensis, redrawn from part of a recording published in[20]).The indicated bandwidth of 20 kHz determines the data rate needed for encoding the signal, whereas the center (peak) frequency of about 84 kHz does not affect the information rate.Adapted from[20].CC BY 3.0.

Figure 5 .
Figure 5. Biomimetic sonar system that integrates a soft-robotic dynamic periphery (noseleaf and pinna) inspired by horseshoe bats with onboard deep-neural-network inference (design rendering by Zhengsheng Lu).Reproduced with permission from Zhengsheng Lu.