An Extensive Application of Model Predictive Control Combined with Policy Search to Multi-agent Agile UAV Flight

Reinforcement Learning (RL) methods can automatically learn complex policies with minimum prior knowledge about the task. Meanwhile, Model Predictive Control (MPC) can achieve excellent control performance with convincing safety and interpretation. Thus, there is a growing amount of research combining the advantages of both RL methods and MPC so that high accuracy and adaptability to highly dynamic environments, or in a word agility, can be achieved during various unmanned aerial vehicles (UAVs) flight tasks. However, current studies mostly solve the control problem for a single UAV. This paper extends such a framework to the problem of controlling a multi-agent system with an MPC controller whose decision variables represented as high-level policies are chosen by policy search. We validate the improved method by flying two drones through a moving gate simultaneously. Experiments in simulation demonstrate that the improved controller can preserve robust, in-time control performance and further avoid collision when there are multiple UAVs, showing a promising aspect for realizing multi-agent agile UAV flight.