Hi, I'm Jen and I'm a digital transformation architect working in digital twins at Booz Allen. Traditional simulation workflows require expensive and distributed computing environments and lack integration for modern-day autonomous systems. They also often require extensive engineering backgrounds to use and have high learning curves, due to the complexity of the user interfaces. To address these challenges, Booz Allen, in collaboration with NVIDIA, has invested in developing a scalable simulation platform that integrates high-fidelity physics and visualization, integration of autonomous systems, and a straightforward user interface. Our Open-World Modeling  consists of three key components: our simulation engine, our user interface, and our metric analytics engine. The first component is our simulation engine: NVIDIA Omniverse. The Omniverse platform provides a high-fidelity visualization and physics platform to build complex simulations. We can write custom Python and C++ code to run these simulations with complex logic. In this scene, we've developed a custom ocean simulation logic to provide physics interaction with boats in the water. We can easily modify the parameters to run millions of unique scenarios, providing ample synthetic data to AI and autonomy algorithms for training and evaluation. We are also able to integrate autonomous systems, such as this autonomous surface vehicle that is running a detection and tracking model as part of its perception stack over ROS 2. With ROS 2 integration, existing autonomy stacks can be seamlessly integrated into Omniverse. The second component is our user interface. Our team of user experience experts conduct user research and iterative user testing to build out seamless and mission-relevant user experiences. We can then rapidly generate frontend web code using code generation software to integrate into our frontend environment. The last component is our  metric analytics engine. As we're evaluating these autonomous systems, we often would like to understand key performance metrics around these systems. In this scenario, we might use the metric data from the perception system to drive selection of a sensor for a specific data capture mission. Booz Allen understands the importance of creating simulations and Digital Twins that integrate real-world autonomous systems within an easy-to-use interface. We've invested internally  into a simulation framework that enables integration of advanced autonomy platforms. What once required significant engineering effort to build is now readily available to the client wanting to test their autonomous systems. By leveraging this workflow, clients will save significant costs in labor  and achieve results faster. 

Hi, I'm Drew Massey, and I'm a chief engineer working in physical AI at Booz Allen. Traditional aerial 3D reconstruction workflows require significant processing time and often focus only on the purely visual elements of a 3D model. They lack the physical characteristics of the world that can inform advanced simulations and digital twins. To address these challenges, Booz Allen,  in collaboration with NVIDIA, has invested in developing a workflow which rapidly and automatically extracts features from video and image data to generate high-fidelity, visually and physically accurate  3D reconstructions from UAV data. Our aerial reconstruction workflow consists of seven key steps. We start with data capture. The techniques that we'll discuss later in the workflow depend on certain aspects of this stage to ensure the highest output quality. To support this, we've developed best practices in oblique capture to ensure that we have adequate and overlapping coverage of an area of interest. This data capture can either be video or image data  taken from an electro-optical camera sensor. Next, we feed this captured data into our feature matching workflow, which utilizes structure-from-motion techniques to produce a sparse and dense point cloud output of overlapping points from the images. From here, we utilize the feature matching outputs to generate Gaussian splats and meshing. For high quality Gaussian splats we leverage the latest NVIDIA fVDB workflow, along with a 3D Gaussian splatting algorithm to generate high quality representations of the environment. For meshing, we've developed modules leveraging Poisson reconstruction and Truncated Signed Distance Field functions to convert the set of points into a mesh. These algorithms build out the 3D geometry of the model, which provides a surface to the object. This mesh can then be used to provide collision for an autonomous asset in simulation. We then perform 3D segmentation of this output using semantic segmentation techniques, to provide an additional layer of metadata that we use to tie physical properties to the visual data. For example, this segmentation model can be used to extract dirt or road material from the data as an image mask. This can then be applied to a roughness texture to the material of the 3D mesh, which will then in turn affect the friction of the material as a ground vehicle drives across it. Finally, we integrate the 3D and segmentation data to build 3D content in our simulation software that will inform AI and autonomy training. This final reconstruction can be used to train, test and evaluate perception models for an autonomous system, such as how the tracking algorithm is being affected by the reflection of a metal barrier, or how the actuation model of a ground robot is affected by loose soil and difficult terrain. Booz Allen understands the importance of creating simulations and digital twins that are both visually and physically accurate. Our automated workflow speeds your  3D reconstruction process and provides an additional layer of metadata and physics. What once took a month or more can now be completed in under an hour. By leveraging our workflow, clients will save significant costs in labor and achieve results faster.