Designing Latent Safety Filters using Pre-Trained Vision Models

2025-09-18

AI Summary

This paper addresses safety-critical challenges in vision-based control systems by proposing the first systematic framework leveraging pre-trained visual representations (PVRs) to construct safety filters. Methodologically: (1) it establishes an interpretable fault-set classification taxonomy; (2) integrates Hamilton–Jacobi reachability analysis with latent world models to dynamically compute safety boundaries; and (3) designs a Q-function–based safety policy switching mechanism, quantifying trade-offs among freezing, fine-tuning, and de novo training. Experiments demonstrate cross-task generalization of safety guarantees across diverse PVRs, uncover correlations between model architecture and safety performance, and introduce an edge-optimized lightweight deployment strategy. The core contribution is the introduction of a PVR-driven safety filtering paradigm that unifies perception uncertainty modeling with control-theoretic safety assurance.

Technology Category

• Computer Vision: Visual Reasoning & Symbolic Representations
• Natural Language Processing: Safety and Robustness
• Reasoning under Uncertainty: Uncertainty Representations

Application Category

• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Ensuring safety of vision-based control systems remains a major challenge hindering their deployment in critical settings. Safety filters have gained increased interest as effective tools for ensuring the safety of classical control systems, but their applications in vision-based control settings have so far been limited. Pre-trained vision models (PVRs) have been shown to be effective perception backbones for control in various robotics domains. In this paper, we are interested in examining their effectiveness when used for designing vision-based safety filters. We use them as backbones for classifiers defining failure sets, for Hamilton-Jacobi (HJ) reachability-based safety filters, and for latent world models. We discuss the trade-offs between training from scratch, fine-tuning, and freezing the PVRs when training the models they are backbones for. We also evaluate whether one of the PVRs is superior across all tasks, evaluate whether learned world models or Q-functions are better for switching decisions to safe policies, and discuss practical considerations for deploying these PVRs on resource-constrained devices.

Problem

• Designing safety filters using pre-trained vision models
• Ensuring safety in vision-based control systems
• Evaluating effectiveness of pre-trained models for safety filters

Innovation

• Pre-trained vision models for safety filters
• Classifiers defining failure sets with PVRs
• Latent world models for safe policies