Lawrence Berkeley National Laboratory

Nuclear material accounting and control (NMAC) has the objective to deter and detect an unauthorized removal of nuclear material from a nuclear facility. In this work, we present an effort to leverage innovations – from the rapidly advancing fields of machine learning, computer vision, and robotics – to automate assaying and inventory tasks. The project is organized along 5 thrusts. First, we build a multi-sensor system that is capable of capturing the relevant information of a scene. Second, we used the system to collect relevant data, mainly at the Nevada National Security Sites, but also a few other sites. Third, we developed techniques to quickly label data and train a neural network, so that the system is able to detect a new type of Nuclear Material Container (NMC) in little time in the field. Fourth, we develop algorithms to predict the location and number of NMCs in a scene and find areas of uncertainty that need to be manually inspected. Fifth, we explored soft robotics and its use for confirmation of presence or absence of NMCs in inaccessible areas. This report discusses work performed in those 5 thrusts in more detail and makes recommendations for future work.