Learning-based Control for Passive Flexible Multirotors (PFMs) with Physical Interaction Abilities

Flexible, deforming aerial vehicles have introduced a paradigm shift in existing research by enabling stable aerial-physical interactions. These vehicles harvest interaction energies to undergo a deformation of the vehicle body and absorb impacts, therefore retaining stability under the action of unknown external forces. The low-level control of such aerial vehicles gets complicated owing to the highly nonlinear and varying dynamics of a passive deforming chassis, coupled with the action of external forces resulting from interactions with the environments. In this work, we propose a morphology-aware Q-learning-based tracking controller that accounts for varying morphology and the various external interaction forces. We train this learning-based controller offline under various case scenarios and then augment an online rollout scheme to improve the tracking performance in case of model-mismatch. Simulation studies are presented that validate the proposed controller in comparison with a conventional disturbance observer approach with improved tracking performance and optimized control effort.