✨ TL;DR
This paper introduces Adaptive Conformal Filtering (ACoFi), which combines learned safety filters with adaptive conformal inference to provide soft safety guarantees for control systems. The method dynamically adjusts switching criteria between nominal and safe policies based on prediction uncertainty, achieving better safety performance than fixed-threshold approaches.
Safety filters are used to ensure control systems remain safe even when nominal policies are unsafe, but traditional synthesis methods face scalability issues with high-dimensional systems. Learning-based safety filters have been proposed as alternatives, but they suffer from inevitable prediction errors that compromise reliability and safety guarantees. The key challenge is how to account for these errors while maintaining safety assurances in real-world applications where the learned models may encounter distribution shifts or make incorrect predictions about action safety.
ACoFi combines Hamilton-Jacobi reachability-based safety filters with adaptive conformal inference to create a dynamic switching mechanism. The method uses conformal prediction to quantify uncertainty in the learned safety filter's predictions by computing a range of possible safety values for the nominal policy's actions. When this uncertainty range suggests potential unsafety, the filter switches from the nominal policy to a learned safe policy. The switching threshold adapts over time based on observed prediction errors, allowing the system to learn from its mistakes and adjust its conservativeness accordingly. This adaptive mechanism is grounded in conformal inference theory, which provides statistical guarantees on the miscoverage rate.
What the paper shows.
ACoFi was evaluated in two environments: a Dubins car simulation and Safety Gymnasium. The method significantly outperformed baseline approaches using fixed switching thresholds, achieving higher learned safety values while incurring fewer safety violations. The improvements were particularly pronounced in out-of-distribution scenarios where the learned models faced conditions different from their training data. The adaptive nature of ACoFi allowed it to adjust to these challenging scenarios more effectively than non-adaptive baselines, demonstrating both better safety performance and less conservative behavior when the nominal policy was actually safe.
The paper provides soft safety guarantees rather than hard safety guarantees, meaning there is a bounded probability of safety violations rather than absolute prevention. The guarantees are asymptotic, requiring sufficient data for the conformal inference bounds to hold reliably. The method's performance depends on the quality of the learned safety filter and the learned safe policy, which may still make errors. The approach requires careful tuning of the user-defined miscoverage parameter to balance safety and performance. Additionally, the evaluation is limited to simulation environments, and real-world deployment may face additional challenges not captured in these settings.
✨ Generated by Claude · Apr 21, 2026 · Read the PDF for authoritative content.