Reinforcement Learning Based Trajectory Planning for Multi-UAV Load Transportation

Abstract

This study introduces a novel trajectory planning approach for the transportation of cable-suspended loads employing three quadrotors, relying on a reinforcement learning (RL) algorithm. The primary objective of this path planning method is to transport the cargo smoothly while avoiding its swing. Within this proposed solution, the value function of the RL is estimated through a feature vector and a parameter vector tailored to the specific problem. The parameter vector undergoes iterative updates via a batch method, subsequently guiding the generation of the desired trajectory through a greedy strategy. Ultimately, this desired trajectory is communicated to the quadrotor controller to ensure precise trajectory tracking. Simulation outcomes demonstrate the capability of the trained parameters to effectively fit the value function.

Julián Estévez

Artificial Intelligence and Robotics Professor

My research interests include distributed robotics, mobile computing and programmable matter.