Title: Ensuring Permissive Safety via Trusted Inference: A Framework for Verifiable Belief-Space Neural Safety Filters in Assured Interactive Robotics

Abstract:

Autonomous systems designed for human interaction face the complex challenge of balancing safety and efficiency amidst human-induced uncertainties, including varying preferences, objectives, skill levels, and cooperative intent. Safety filters have emerged as a prevalent solution in interactive robotics, offering a modular architecture that decouples safety constraints from performance metrics. This separation enables robots to navigate around humans safely while maintaining minimal disruption to task efficiency. Although conventional safety filters generally function within physical space and often overlook the robot's capacity for online learning and adaptation, the recently introduced belief-space safety filter (BeliefSF) addresses this limitation. BeliefSF evaluates safety within a closed-loop framework, utilizing runtime inference to actively diminish uncertainty during operation, which in turn lowers the conservativeness inherent in traditional filtering methods.

Nevertheless, establishing formal safety guarantees for robots employing BeliefSF is difficult, primarily due to errors arising from runtime inference and the neural approximation techniques necessary to manage the high dimensionality of belief spaces. To address this, we present an algorithmic strategy to certify the high-probability safety of BeliefSF by integrating conformal prediction, with explicit consideration for the reliability of the robot’s runtime inference module. Our approach capitalizes on the structural properties of belief-space safety filtering by concentrating verification efforts on regions where inference is anticipated to be dependable. This method maintains the simplicity and sample complexity characteristic of standard conformal prediction while enabling the certification of a significantly less conservative safety filter. In a simulated benchmark involving human-vehicle interactions, our results demonstrate that the proposed approach verifies a markedly more permissive belief-space safety filter compared to a baseline using standard conformal prediction.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC