Transformer-based data-driven reinforcement learning for collision-free integrated planning and control of multiple UAVs

This paper studies the challenging collision-free planning and control problem for multiple Unmanned Aerial Vehicles (UAVs) with complex dynamics. A data-driven reinforcement learning framework is constructed to address this challenge by combining the transformer-based learned dynamics and iterative Linear Quadratic Regulator (iLQR) optimization. First, each UAV employs an independent transformer network, Multi-Head Self-Attention (MHSA), and residual connections to model local dynamics from online collected data. This approach enables efficient Jacobian computations via parallelization by exploiting the inherent block-diagonal structure in the decoupled dynamics. Then, to avoid inter-UAV and UAV-obstacle collision in the cooperative flight, logarithmic barrier functions are incorporated into the cost function of iLQR. The block-diagonal approximation of the Hessian is employed to overcome the coupling induced by the barrier terms and preserve the computational tractability during the backward pass. Specifically, the proposed framework possesses robust collision avoidance capabilities in solving the multi-UAV planning and control problem. Finally, simulation results demonstrate the effectiveness and superiority in convergence speed and accuracy of the proposed framework.