Trading-Off Safety with Agility Using Deep Pose Error Estimation and Reinforcement Learning for Perception-Driven UAV Motion Planning

Navigation and planning for unmanned aerial vehicles (UAVs) based on visual-inertial sensors has been a popular research area in recent years. However, most visual sensors are prone to high error rates when exposed to disturbances such as excessive brightness and blur, which can lead to catastrophic performance drops in perception and motion planning systems. This study proposes a novel framework to address the coupled perception-planning problem in high-risk environments. This achieved by developing algorithms that can automatically adjust the agility of the UAV maneuvers based on the predicted error rate of the pose estimation system. The fundamental idea behind our work is to demonstrate that highly agile maneuvers become infeasible to execute when visual measurements are noisy. Thus, agility should be traded-off with safety to enable efficient risk management. Our study focuses on navigating a quadcopter through a sequence of gates on an unknown map, and we rely on existing deep learning methods for visual gate-pose estimation. In addition, we develop an architecture for estimating the pose error under high disturbance visual inputs. We use the estimated pose errors to train a reinforcement learning agent to tune the parameters of the motion planning algorithm to safely navigate the environment while minimizing the track completion time. Simulation results demonstrate that our proposed approach yields significantly fewer crashes and higher track completion rates compared to approaches that do not utilize reinforcement learning.