Find in Library

This paper presents distributed training approach for a nonlinear and heterogeneous multi-UAV system to solve a safe and optimal formation control problem. The objective of control is to ensure safety while achieving optimal performance. In this regard, the position and attitude controllers are considered in series. First, the optimal formation control design is defined as the optimal performance in position control and is modeled by the cost function. In this article, with the integration of cost functions and local control barrier functions (CBFs), a novel distributed optimization problems are introduced. Existing the local CBF in the augmented cost function ensures the safety of the position control, and as a result, collisions do not occur along the path of UAVs. The proposed method considers the safe and optimal position controllers by solving unconstrained optimization problems instead of constrained ones. In the next stage, the reference attitudes are driven by virtual position control. The attitude tracking optimal control is considered the optimal performance in the attitude control, and the related cost function models it. Finally, the stability and safety of the proposed controllers are proven. These optimal and safe policies are obtained sequentially using off-policy multi-agent reinforcement learning (MARL) algorithms which do not require knowledge of UAVs' dynamics. The proposed algorithms are validated by simulating the formation control problem of 6 UAVs with collision avoidance constraints.