Automated Discovery via VPP Experimentation, Research, Development, and Technologies (ADVERDT)

To fulfill the critical needs of DTRA, we propose a comprehensive technical approach to assess the feasibility of repurposing VPP computer vision models to deliver precise, reliable, and robust solutions for CBRN signature detection that surpass the performance of state-of-the-art computer vision approaches and models. We also assess the feasibility of using VPP computer vision and CGI to create ultra-realistic immersive training environments. Our Phase I Technical Approach leverages scientific research and experimentation to thoroughly examine the feasibility of adapting, scaling, and optimizing VPP models to align with DTRA's priorities and objectives. These objectives encompass crucial elements such as motion, anomaly, and change detection, as well as addressing complex challenges like image and video tampering, deepfake detection, and zero, one, and few-shot learning of classes and objects. Additionally, our assessment extends to exploring the practicality of using VPP approaches and models seamlessly render real-world objects into immersive environments tailored to the unique planning and training needs of operators and inspectors. For each Priority Area and Objective identified by DTRA, our approach evaluates the impact of data quality and availability on model accuracy, precision, and robustness. We explore the adaptability of VPP models to meet these specific objectives and assess the feasibility of scaling these models to optimize execution without compromising predictive accuracy. The Phase I Technical Feasibility study will inform the development of a Technical Roadmap to guide the execution of a Phase II Test Plan that defines specific test cases, success criteria, and objectives for applying VPP-derived methods and models to one (threshold) or two (objective) priority areas.