Autonomous Robotic Inspection based on Active Vision and Deep Reinforcement Learning

Abstract

Over the years, various methods based on computer vision have been proposed for the problem of damage detection and segmentation. Recently, remarkable progress has been made in this field owing to the emergence of deep neural networks. However, there is a main assumption among these works that the data collection (e.g., taking photos) is usually carried out by human inspectors, so there is little occlusion or bad lighting conditions. With that being said, the uncertainties that could occur in data collection are handled manually to ensure the dataset is clean and free of confusing or occluded damages. This assumption limits the applicability of autonomous robotic inspection to real-world settings due to the uncertainties in data collection and data interpretation. To bridge this gap, this study integrates the concept of active perception into damage detection and proposes a framework based on the Partially Observable Markov Decision Process (POMDP) and Deep Reinforcement Learning (DRL). The proposed framework facilitates the learning process for robotic agents to explore the 3D environment and intelligently select informative viewpoints to reduce uncertainty and minimize confusion, which leads to more reliable decision-making. Besides uncertainty reduction, the DRL agent can also inspect the workspace more efficiently compared with traditional raster scanning. The trained DRL agent is evaluated for the autonomous assessment of cracks on metallic surfaces. Results show that the agent equipped with the active perception module outperforms the raster scanning inspection by 57% in terms of crack IoU. In addition, the DRL agent can reduce the total inspection time by two times while the prediction accuracy is on par with the raster scanning.