UAV planners are safety-critical software components whose behavior depends on high-dimensional, continuously tunable parameters. While existing fuzzing approaches focus on discovering failures, configurations deep inside failure regions are often diagnostically uninformative.Once the planner moves far beyond the safety boundary, parameter sensitivity diminishes, and inter-parameter dependencies become obscured.We refer to this phenomenon as diagnostic saturation and observe that configurations near the safe-unsafe transition preserve informative parameter sensitivity and interaction structure, making them more suitable for root-cause analysis. To this end, we propose Planner-Fuzz, a boundary-oriented fuzzing framework. It steers search toward a narrow band around the safety boundary while preventing convergence to a single configuration to enable the exploration of diverse boundary behaviors. Using the collected boundary configuration, Planner-Fuzz performs three-level diagnosis through sensitivity analysis, interaction mining, and counterfactual validation to identify and validate the parameters responsible for failure behavior. We evaluate the approach on three open-source UAV planners under four anomaly types. The results show that Planner-Fuzz discovers more diverse and diagnostically useful boundary configurations than failure-driven baselines, provides stronger signals for identifying influential parameters and interactions, and yields stronger intervention-based recovery evidence. These findings suggest that fuzzing for safety-critical planners should go beyond failure discovery and explicitly support diagnosis.

@inproceedings{paperbot12455,
  title = {Planner-Fuzz: Overcoming Diagnostic Saturation in UAV Planners via Boundary-Oriented Fuzzing},
  author = {Rongbo Chen and Chengcheng Zhao and Peng Cheng and Jiming Chen and Taegyu Kim},
  booktitle = {ASE},
  year = {2026},
  doi = {10.1145/3832783.3834389}
}