רון שטרן

Automated planning algorithms rely on having a sufficiently accurate domain model to solve planning problems, particularly in complex environments with numeric effects. In real-world settings, however, domain models can become misaligned with the environment over time, i.e., no longer accurately reflect the environment. Even a minor drift in numeric effects can lead to faulty plans and execution failures. In this work, we address the problem of repairing numeric planning domain models by proposing FELIX, a learning-based repair mechanism that restores planning effectiveness by identifying and correcting outdated numeric effects. Rather than re-learning the full environment dynamics from scratch, our approach updates only the affected components of the domain model, enabling efficient adaptation with limited data. FELIX restores planning performance across diverse numeric domains, even under complex model drift. It uniquely handles both nonlinear effect shifts and structural signature changes while maintaining polynomial repair overhead under bounded parameter-type assumptions.