Experimental results for autonomous model-predictive trajectory planning tuned with machine learning

This paper presents experimental results of a high-level trajectory planning algorithm for autonomous quadrotors based on Model Predictive Control (MPC) tuned with machine learning. Time-varying planar inequality constraints are used to avoid obstacles. The nonlinear plant dynamics are linearized around a hover condition. Learning Automata is used to select the relative weights of the objective function and compensate for nonlinearities lost during this linearization. The proposed technique successfully guides a quadcopter to a target while avoiding a spherical obstacle placed in its path. These results demonstrate the potential application for MPC-based techniques in unmanned aerial vehicle operations that involve obstacles. Furthermore, they demonstrate that machine learning can be used to tune parameters in an MPC formulation.