Leveraging Temporal Dynamics with Neural-Inspired Sensing and Control

Abstract

Flying insects are known for their fast and robust control while being challenged with sensory delays, an unsteady environment and by having limited computation power. One important component of this robust control is the sensory feedback from arrays of mechanoreceptors found on wings and wing-derived halteres. By combining structural simulation with experimentally derived neural processing models we gain insight into mechanisms involved in detecting body rotation by mechanosensory oscillating appendages. I found that it is the combination of the temporal encoding of strain by mechanoreceptors with the spatial layout of the sensors on the wing that allows for the detection of minute rotation-induced differences in wing deformation. Although several studies have presented analytical models of haltere deformation, a high fidelity Finite Element Analysis (FEA) revealed novel deformation modes resulting from haltere asymmetry. Using a neuronal spiking model on the strain from the FEA simulations, we found spike timing along the circumference of the haltere base changed with body rotation. The timing change was larger than the experimentally-observed timing variability of the individual mechanosensors at all but the top and bottom of the haltere base. This gives credence to the hypothesis of timing-based detection and encoding of rotation, in addition to the recruitment based detection commonly described in the literature. The importance of timing in mechanosensation in insect flight led to the investigation of a timing-based feedforward controller that I tested on a the partially denied inverted pendulum. Using this timing-based feedforward controller, a close-to-optimal controller could be learned in much fewer trials than a brute force search. This neural-inspired controller holds promise for engineered systems where the number of trials is limited and state measurements are denied in parts of it's state space.