Robot Path Planning Utilizing Object Recognition for Inspection of Nuclear Material Containers

摘要

Recent advances in machine learning (ML) based object recognition have enabled robotic systems to autonomously navigate to and investigate sources of radiation such as nuclear material containers. This paper demonstrates an autonomous system using a quadruped robotic platform. It autonomously surveys a nuclear facility and iteratively identifies nuclear material containers to locate target objects and investigates them by performing radiological measurements, which can assist in nuclear safeguards inspections. This capability is enabled through a suite of sensors called Localization and Mapping Platform mounted on the robot that maps a local scene and fuses the spectroscopic radiation data with camera imagery and a 3D representation of the environment. This paper demonstrates automatic navigation to a target object within the scene in minutes using a Robot Operating System interface, automatically detect target objects in the camera imagery using an instance segmentation model, and quantitatively assess the radiation signatures emitted from the identified objects. The instance segmentation model was trained on images of nuclear material containers, which were labeled using a novel semi-automatic annotation method, allowing for hundreds of high-quality images to be labeled within 15 minutes. After training, the object detection results were fused with LiDAR data to create a consistent representation of the observed space and the locations of the detected object. A measurement campaign was conducted at the Nevada National Security Site where the robot's path planning and object recognition capabilities were tested. The robot's path-planning algorithm received the 3D coordinates of objects identified through ML-based object detection and autonomously navigated to those targets, performing radiation detection measurements that could support compliance determinations for nuclear safeguards protocols by quantitatively reconstructing the gamma-ray activity emitted from the detected objects.