Integrative Mechanosensing in Diptera: iMechFly

The development of agile, adaptable and robust engineered systems continues to be accelerated by insights from nature. To control fast movement, animals must rapidly and effectively estimate the instantaneous state of motion and loading of their body-parts using sensors. The same is true for engineered systems such as autonomous vehicles and wearable technology, yet they typically use much more computationally-intensive neural networks. Artificial neural networks are famously bio-inspired and have revolutionised data processing but they continue to operate on von Neumann-style processors, take inputs from traditional sensors, and often fail in unfamiliar situations outside those experienced during training. This is at odds with evolved locomotor control strategies in animals, which have fundamentally different approaches and exhibit superior performance across a range of tasks, including dealing with novel, sparse, noisy and incomplete data. One theory for why animal sensorimotor performance excels, is the co-evolution of mechanosensory apparatus with the architecture it observes. This enables passive, physical filtering of the potential broadband information, passing through only salient features to be encoded by the peripheral nervous system: this is the essence of morphological computing. A key challenge for biologists and engineers alike is to understand how sensorimotor control is achieved, leading us to hunt for tractable animal models that can be interrogated comprehensively. The true flies (Diptera, including the genetic model species Drosophila melanogaster) have been particularly useful in this respect, with an ever more complete knowledge of how the compound eyes and the halteres (vestigial hindwings instrumented with strain sensors that detect body rotations via inertial forces) operate, with mechanosensation having a speed advantage over visual processing. The wings are richly sensorised in a homologous manner to the halteres and experience a complex combination of inertial and aerodynamic forces. However, they are less studied, leaving an incomplete understanding of sensorimotor flight control. Our aim here is to create a general model of wing mechanosensing that reveals the principles of co-evolved sensor distribution and morphology. We will measure multi-scale body and deforming wing kinematics of eight fly species with diverse wing shapes, wingbeat frequencies, and flight behaviours. We will test hypotheses of species-specific sensor placement patterns, to determine whether locations are matched to the aeroelastic strains experienced by each species across their typical flight parameter space. This will reveal whether the architecture is filtering sensory cues prior to encoding—such that perception is tuned to useful information only—thus reducing the computational burden and demonstrating morphological computing (also known as embodied intelligence). We will integrate our findings with published electrophysiological responses of strain sensors to test whether sensitivity to physical cues matches the signals they experience. Finally, we will incorporate a scaling analysis of wing architecture and wingbeat frequency to maximise the applicability of the general principles we discover to embodied control systems more widely. The large dataset of deforming wing kinematics, from multiple fly species, will be unprecedented in scope and detail, forming a valuable open source dataset. Our results will be of significant interest to biologists and engineers, with application to sensory physiology, arthropod evolution, functional morphology, neuromorphic computing, sparse sensing, and adaptive control for autonomous vehicles and robotics. This is curiosity-driven, data-intensive, frontier bioscience addressing fundamental questions in biology. It sits within the BBSRC priority area Frontier bioscience: understanding the rules of life, with clear application to Transformative technology.