Characterization of Shape Memory Alloy Behavior and Position Control Using Reinforcement Learning

Intelligent, autonomous, shape-controllable structures based on Shape Memory Alloy materials have the potential to provide advanced aircraft and spacecraft systems with the ability to morph, or change their shape for the purpose of optimizing performance. An important aspect of this capability is control of the shape modifications themselves, which benefits from accurate models of the voltage/current-force/deformation relationships of the Shape Memory Alloy materials. These models are typically developed from a constitutive relation for the Shape Memory Alloy behavior which is then integrated into a structural model. The approach presented in the paper does not need a constitutive model, but uses Reinforcement Learning to directly learn an input-output mapping characterization in realtime. The control voltage outputs required to produce a certain shape change are determined using a simulated Shape Memory Alloy test rig, which is integrated with a Qlearning Reinforcement Learning software module. The goal of this research is to use this characterization to determine an optimal control policy when changing the shape of a Shape Memory Alloy to a specified length. The results of this research are expected to aid in characterizing the effectiveness of this type of advanced control mechanism in intelligent systems, and further research in the modeling and control of morphing air and space vehicles. Results presented in this paper show that a hyperbolic tangent function can successfully be used to simulate Shape Memory Alloy hysteresis behavior and that Reinforcement Learning is a viable tool in characterizing Shape Memory Alloy behavior.