Go to 2024 Workshop MAPoDeL homepageMAC: Morphology-Aware Control Design with Deep Reinforcement Learning for Navigation of Squeezable Unmanned Aerial Vehicles
Traditional methods optimize robot morphology independently using evolutionary algorithms, where fitness functions are applied to evaluate each genotype. Control strategies are then designed based on this predefined morphology. While the concept of co-design, considering multiple aspects simultaneously, is not new, it is often impractical due to the time-consuming process of manufacturing new morphology. A common solution to this limitation is restricting the design space to morphology with similar topologies. In this extended abstract, we investigate a squeezable \gls{uav} as the controlled plant and propose a novel \gls{mac} method leveraging \gls{drl} to solve a vision-based navigation problem under extreme settings in this study. Our approach integrates the agent’s morphology directly into the learning process of the control policy, enabling fast morphology-policy co-design. To simplify the control problem, we restrict the quadrotor configuration that can be transformed only between X and H via a single squeezing angle ($\xi$). The quadrotor's body shape can be horizontally reduced to 52.4 percent of its original shape. Simulation results show that the navigation policy trained with an understanding of morphology is more effective.