Award

State of the art AI operational planning tools predominantly use hand crafted planning models, which can be expensive to develop and maintain. We addressed this shortcoming with our recent ONR SBIR Phase II effort CILEMP, Continuous Interactive Learners for Misson Planning, where we demonstrated the ability to learn planning models from synthetically generated structured data. However, during our transition efforts, we identified the need for extracting and learning planning models from unstructured sources as a capability gap. To enable widespread adoption of AI plan model learning tools, and to increase the scale of autonomy, we propose to extend CILEMP with FLAG (CILEMP-FLAG), Fleet Autonomy with Generative AI. Specifically, we will investigate a novel distributed architecture for multiple autonomous agents that collaboratively decide, plan, and act using decision transformers and LLMs. To this end, we will systematically investigate and address the performance limitations of LLMs; namely, non-deterministic output, over generation or hallucinations, prohibitively large computes, and expensive solution validation in complex tasks such as planning. We will use a divide and conquer approach to reduce problem complexity and orchestrate multiple lightweight LLMs to reduce the compute requirements, together with simulated validation and relevant constraint checkers. This novel framework will extend the CILEMP learning framework and enable us to not only reach unprecedented levels of accuracy for model extraction and narrative generation but also ameliorate the limitations of LLMs.  We will demonstrate and validate the CILEMP-FLAG plan model extraction from unstructured text and multimedia narrative generation in Navy relevant missions such as ISR and Mission Support Logistics. We will deliver the CILEMP-FLAG software system at TRL-6.