Model-Free Guidance Method for Drones in Complex Environments Using Direct Policy Exploration and Optimization

Abstract

In the past few decades, drones have become lighter, with longer hang times, and exhibit more agile performance. To maximize their capabilities during flights in complex environments, researchers have proposed various model-based perception, planning, and control methods aimed at decomposing the problem into modules and collaboratively accomplishing the task in a sequential manner. However, in practical environments, it is extremely difficult to model both the drones and their environments, with very few existing model-based methods. In this study, we propose a novel model-free reinforcement-learning-based method that can learn the optimal planning and control policy from experienced flight data. During the training phase, the policy considers the complete state of the drones and environmental information as inputs. It then self-optimizes based on a predefined reward function. In practical implementations, the policy takes inputs from onboard and external sensors and outputs optimal control commands to low-level velocity controllers in an end-to-end manner. By capitalizing on this property, the planning and control policy can be improved without the need for an accurate system model and can drive drones to traverse complex environments at high speeds. The policy was trained and tested in a simulator, as well as in real-world flight experiments, demonstrating its practical applicability. The results show that this model-free method can learn to fly effectively and that it holds great potential to handle different tasks and environments.

Design and Control of Autonomous Flying Excavator

“…Additionally, an optimization method is introduced to determine the optimal flight path, enabling the accurate recognition of as many objects as possible while minimizing time costs. In [17], a new model-free reinforcement-learning method is introduced, offering the capacity to learn the optimal planning and control policy from flight data. In the training phase, this method considers the complete drone state and environmental data as inputs.…”