A brain for a batbot: Combining deep learning and biomimetic robots to understand and replicate bat biosonar

Sonar-based navigation in complex natural environments—whether underwater or in-air—poses major scientific and technological challenges. A key factor for these problems is the unpredictable nature of echoes that are superpositions of contributions from many reflectors (“clutter”). However, many bat species thrive in dense vegetation and hence demonstrate every night that such echoes can convey copious amounts of sensory information. Deep learning (DL) provides a fresh look at these difficult problems with the possibility of taking performance to new levels. A critical advantage of DL is its superior ability to discover informative features. However, DL requires training data sets that are typically much larger than what can be obtained in experiments with behaving animals. Biomimetic robots that can reproduce the behaviors of bats are a good way to obtain large-enough data sets of real-world echoes that have been recorded under controlled conditions. Examples for this approach are the exploration of biosonar landmarks and passageway finding in natural environments. For even larger data sets with better control over the underlying parameters, data augmentation methods based on generative networks can be used. An important remaining challenges is understanding the features selected by the DL networks at the signal and physical levels.