From Demonstrations to Safe Deployment: Path-Consistent Safety Filtering for Diffusion Policies

2025-11-09

AI Summary

Diffusion policies (DPs) achieve strong performance in complex manipulation tasks but lack formal safety guarantees, relying instead on external safety mechanisms that often induce distributional shift and degrade task performance. To address this, we propose Path-Consistent Safety Filtering (PACS), a trajectory-level safety framework that integrates consistency-aware braking with set-based reachability analysis to enable real-time, environment-aware safety verification during action generation—ensuring all safety interventions remain strictly within the DP’s training distribution. PACS is the first safety framework for diffusion policies that provides formal safety guarantees without compromising original task success rates. Experiments across simulation and three real-world human-robot interaction tasks demonstrate that PACS improves task success rates by up to 68%, significantly outperforming conventional reactive safety methods.

Technology Category

• Natural Language Processing: Safety and Robustness
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Philosophy and Ethics of AI: Safety, Robustness & Trustworthiness

Application Category

• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• Security and Privacy: Data transparency and provenance
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Diffusion policies (DPs) achieve state-of-the-art performance on complex manipulation tasks by learning from large-scale demonstration datasets, often spanning multiple embodiments and environments. However, they cannot guarantee safe behavior, so external safety mechanisms are needed. These, however, alter actions in ways unseen during training, causing unpredictable behavior and performance degradation. To address these problems, we propose path-consistent safety filtering (PACS) for DPs. Our approach performs path-consistent braking on a trajectory computed from the sequence of generated actions. In this way, we keep execution consistent with the policy's training distribution, maintaining the learned, task-completing behavior. To enable a real-time deployment and handle uncertainties, we verify safety using set-based reachability analysis. Our experimental evaluation in simulation and on three challenging real-world human-robot interaction tasks shows that PACS (a) provides formal safety guarantees in dynamic environments, (b) preserves task success rates, and (c) outperforms reactive safety approaches, such as control barrier functions, by up to 68% in terms of task success. Videos are available at our project website: https://tum-lsy.github.io/pacs/.

Problem

• Ensuring safety guarantees for diffusion policies in dynamic environments
• Maintaining task success rates while implementing safety mechanisms
• Preventing performance degradation from external safety interventions

Innovation

• Path-consistent safety filtering for diffusion policies
• Set-based reachability analysis for real-time safety verification
• Trajectory braking maintaining policy training distribution consistency