Selective Agentic Recovery for UAV Autonomy with a Persistent Mission Runtime

The paper introduces the Persistent Mission Runtime (PMR), a UAV recovery framework that selectively invokes an external agentic reasoner for high-level decision-making during mission execution. PMR utilizes a learned Cognitive Value of Invocation (learned-CVI) to determine when to engage the reasoner, significantly enhancing success rates from 5.0% to 95.0% on a 400-run Gazebo/PX4 benchmark, while reducing remote agent calls by 16.7% and logged tokens by 29.2%. This approach allows UAVs to maintain local autonomy, optimizing resource usage and improving mission efficiency, which is critical for practitioners developing robust UAV systems.