How to train AI for the real world using a synthetic environment

13 May 2027
New Horizons Keynote Theatre
Outpacing Autonomous and Uncrewed System Threats

To get ahead of emerging threats in modern defense operations, warfighters require rapid, reliable AI detection capabilities. While AI can be trained to do almost anything, its success hinges entirely on data. To build a working model, you need data from the right context, with perfect labeling, varied lighting conditions, and legal clearance. Gathering this is a massive hurdle for any industry, but in defense, it is exceptionally difficult. A robust AI detection model doesn't just need a single image of a military object on a road or in a sky. It needs thousands of variations, captured by different sensors across countless environments. Collecting these real-world datasets is notoriously expensive, slow, and cumbersome. But what if the solution is to build the data rather than collect it? Civilian dominated datasets leave defense AI unequipped for specific military domains. Using a synthetic environment which is built for military training allows the creation and training of AI using domain specific datasets. This is critically important because the AI accuracy and trust is only as good as the data AI models are trained with. By leveraging fully immersive 3D synthetic environments, developers can create custom worlds to extract exactly the data they need to train domain specific AI. The question then becomes: can we successfully use synthetic scenes to train AI for the real world? By bypassing the massive hurdle of real-world data collection, synthetic environments allow defense organizations to iterate models rapidly and deliver AI-enabled detection capabilities to the warfighter at pace.

Chairperson
Benjamin Bell
Benjamin Bell, General Manager - Advanced Capabilities Engineering, LLC
Speakers
Natalija Dierenbach
Natalija Dierenbach, Solutions Engineer - BAE Systems OneArc
Oliver Arup
Oliver Arup, CTO/CPO - BAE Systems OneArc