Uncertainty-Aware Autonomous Robotic Inspection Based on Active Vision and Deep Reinforcement Learning

Abstract

This paper introduces an uncertainty-aware active vision framework, UADS-DRL, for autonomous robotic damage inspection. Unlike conventional methods that rely on passive raster scanning or prior learning-based agents, UADS-DRL formulates the task as a partially observable Markov decision process, enabling dynamic viewpoint selection guided by segmentation uncertainty. The agent leverages uncertainty cues to fuse predictions from multiple time steps, thereby achieving robust segmentation. Experimental results on photorealistic metallic surface inspection show that UADS-DRL achieves more than 2 times improvement in damage mIoU and over 50% reduction in inspection time compared to dense overlap raster scanning. It also surpasses the earlier ADS-DRL baseline by 20% in damage IoU with comparable inspection time. Given similar IoU performance, the proposed UADS-DRL agent can cut the inspection time by more than half compared with the previous learning baseline. These results underscore the effectiveness of integrating uncertainty estimation for accurate and time-efficient robotic inspection.