Designing Latent Safety Filters Using Pre-Trained Vision Models
Ihab tabbara, yuxuan yang, Ahmad Hamzeh, Maxwell Astafyev, Hussein Sibai
AI summary
Problem
Vision-based control systems lack reliable safety guarantees in complex environments because traditional safety filters do not scale to high-dimensional visual inputs. Existing methods lack a systematic evaluation of how pre-trained vision representations can be adapted for safety-critical filtering.
Approach
The authors systematically evaluate five state-of-the-art pre-trained vision models as backbones for latent failure classifiers and Hamilton-Jacobi reachability safety filters, comparing frozen, fine-tuned, and from-scratch training regimes across multiple simulated robotics tasks.
Key results
- Fine-tuning PVR backbones reduces safety violations by 73.7% compared to training from scratch
- DINOv2 consistently delivers robust safety filter performance across all evaluated tasks
- World model-based safety assessment outperforms Q-function critics when the backbone captures dynamics well
- PVR-based latent safety filters maintain computational efficiency suitable for real-world deployment
Why it matters
Demonstrates that leveraging foundation vision models provides a scalable, data-efficient pathway for deploying safe vision-based control in robotics.
Abstract
Ensuring safety of vision-based control systems re- mains 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 representations (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 value functions, and for latent world models. We discuss the trade-offs between training from scratch, fine-tuning the PVRs, 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. Our experiments show that compared to training representations from scratch, using PVRs as perception backbones for vision-based safety filters can reduce violation rates by 12.2%, and fine-tuning PVRs to the task can reduce them by 73.7%, while maintaining or improving task performance. Code is available at https: //github.com/tabz23/Latent-Safety-Filters.