Safe Control using Learned Safety Filters and Adaptive Conformal Inference

2026-04-20 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Adaptive Conformal Filtering (ACoFi) is a novel method that integrates learned Hamilton-Jacobi reachability-based safety filters with adaptive conformal inference to enhance the safety of control systems. This approach addresses the scalability limitations of traditional safety filter synthesis, particularly in high-dimensional state and control spaces, while mitigating concerns about prediction errors inherent in learning-based models. ACoFi dynamically adjusts its switching criteria based on observed errors in safety predictions, using the range of possible safety values of the nominal policy's output to quantify uncertainty. The filter switches from the nominal to the learned safe policy when this range indicates potential unsafety. ACoFi guarantees that the rate of incorrectly quantifying uncertainty in predicted nominal policy safety is asymptotically upper bounded by a user-defined parameter, offering a soft safety guarantee. Empirical evaluations in a Dubins car simulation and a Safety Gymnasium environment show ACoFi significantly outperforms baseline methods with fixed switching thresholds, achieving higher learned safety values and fewer safety violations, especially in out-of-distribution scenarios.

Key takeaway

For research scientists developing control systems with learned safety filters, ACoFi offers a robust approach to enhance reliability and reduce safety violations. By dynamically adjusting switching criteria based on prediction uncertainty, you can achieve a quantifiable soft safety guarantee, particularly beneficial in complex, high-dimensional, or out-of-distribution environments. Consider integrating ACoFi to improve system performance and safety compared to traditional fixed-threshold methods.

Key insights

ACoFi combines learned safety filters with adaptive conformal inference for robust, scalable control system safety.

Principles

• Dynamic adjustment improves safety.
• Quantify uncertainty for reliability.
• Soft guarantees are achievable.

Method

ACoFi dynamically adjusts switching criteria based on observed prediction errors, using the range of possible safety values to quantify uncertainty and switch to a learned safe policy when needed.

In practice

• Apply ACoFi to high-dimensional control.
• Use adaptive switching over fixed thresholds.
• Improve safety in OOD scenarios.

Topics

• Adaptive Conformal Filtering
• Learned Safety Filters
• Hamilton-Jacobi Reachability
• Conformal Inference
• Control Systems Safety

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• Permissive Safety Through Trusted Inference: Verifiable Belief-Space Neural Safety Filters for Assured Interactive Robotics — A new method certifies high-probability safety for belief-space neural safety filters in interactive robotics, enabling less conservative operation via conformal prediction.
• Adaptive Confidence Regularization for Multimodal Failure Detection — ACR detects multimodal failures by penalizing confidence degradation and synthesizing failure-aware training examples.
• A Subjective Logic-based method for runtime confidence updates in safety arguments — A Subjective Logic method dynamically updates safety case confidence using runtime Safety Performance Indicators.
• Efficient Online Conformal Selection with Limited Feedback — The ACI update rule ensures valid and efficient conformal selection even with limited bandit feedback.
• Responsible Federated LLMs via Safety Filtering and Constitutional AI — Integrating safety filters and Constitutional AI significantly improves FedLLM safety against harmful content.
• On the Impossibility of Separating Intelligence from Judgment: The Computational Intractability of Filtering for AI Alignment — External filtering is computationally intractable for aligning LLMs, implying intelligence and judgment are inseparable.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.