Several of the next generation nuclear reactors being considered will require autonomous maintenance capabilities, due to the hostile radiation and thermal environments that exclude human operators from supporting equipment in close proximity to the reactors. Research is ongoing to use virtual training environments to train robots to perform maintenance and emergency response operations. Since future designs are still under development, there is no testbed for physical experiments. To overcome this obstacle, we have developed a generalized methodology to train robots in a virtual environment. In this regard, the virtual environment can be created from either 3-D mapping of existing physical components or from components that only exist as engineering specifications, as long as their dimensions and behaviors can be represented in the virtual world. Our approach creates an integrated software and algorithm architecture with the ability to “teach” automated systems successful tasks from simulated data sets with minimal human oversight. In this work, we develop a digital-twin system for training and real-time control of a robotic arm. The digital twin comprises a virtual training environment and a real-time control interface. The virtual training environment allows simulation of maintenance tasks under diverse domain randomizations. After training, the physical robot is controlled while monitoring anticipated behavior of the digital twin in real time. This approach increases operator confidence and can decrease risk compared to deployment of a model without real-time supervision of the digital-twin alongside the physical machine. It is clear that this approach has extensive applications to terrestrial and space-based tasks. Autonomous systems including robots, will support future maintenance and security protocols and this method provides one possible approach to address future uncertainties or responses to accidents. This technology is in its infancy but expected to mature rapidly and undertake complex and dynamic tasks using reinforcement learning derived algorithms.