Learning Variable Whole-Body Control for Agile Aerial Manipulation in Strong Winds

Abstract

Aerial manipulation provides an effective alternative to human labor in high-risk outdoor situations. Complex and variable environments demand the system to respond quickly with minimal latency to external disturbances. To address this challenge, we propose a learning-based variable whole-body model predictive controller designed to improve the adaptability and agility of the system through robotic arm-assisted motion. Given the limited onboard computing power, this low-level whole-body model predictive controller enhances computational efficiency without sacrificing accuracy by linearizing the highly coupled dynamics model and updating the linearized parameters in real-time. By incorporating updates of the disturbance values predicted by the Gaussian process into the linear model, the whole-body controller can swiftly react to perturbations. Additionally, it can employ robotic arm motions to perform agile maneuvers and counter disturbances, rather than merely adjusting the quadrotor's rotational movements. To further enhance agility and robustness, we train a high-level policy search using episode-based policy search and gradient descent techniques. For specific tasks and scenarios, this policy search can train a deep neural network to identify optimal decision variables that account for various wind disturbances for the low-level controller. We have carried out disturbance rejection and flip experiments on the aerial manipulation system in the wind tunnel, which demonstrate that the controller can operate stably and effectively under strong disturbance.