From Demonstrations to Safe Deployment: Path-Consistent Safety Filtering for Diffusion Policies
Ralf Römer, Julian Balletshofer, Jakob Thumm, Marco Pavone, Angela P. Schoellig, Matthias Althoff
AI summary
Problem
Diffusion policies achieve state-of-the-art performance in complex manipulation but lack safety guarantees, and existing reactive safety filters often drive robots into out-of-distribution states, causing unpredictable behavior and significant performance degradation.
Approach
PACS converts the policy’s predicted action chunks into a smooth intended trajectory and applies high-frequency braking along that exact path, using set-based reachability analysis to verify safety constraints in real time.
Key results
- Provides formal safety guarantees in dynamic environments without compromising task success
- Outperforms reactive safety filters by up to 68% in task success rate during simulation
- Achieves 37% higher task success rates than reactive filters in real-world hardware experiments
- Computing trajectories from full action chunks improves success rates by 28% over sequential single-action processing
Why it matters
Enables reliable deployment of state-of-the-art generative robot policies in safety-critical human-robot interaction scenarios where both formal safety guarantees and high task performance are required.
Abstract
Diffusion policies (DPs) achieve state-of-the-art performance on complex manipulation tasks by learning from large-scale demonstration datasets, often spanning multiple em- bodiments and environments. However, they cannot guarantee safe behavior, requiring external safety mechanisms. These, however, alter actions in ways unseen during training, causing unpredictable behavior and performance degradation. To ad- dress these problems, we propose path-consistent safety filtering (PACS) for DPs. Our approach performs path-consistent brak- ing on a trajectory computed from the sequence of generated actions. In this way, we keep the execution consistent with the training distribution of the policy, maintaining the learned, task-completing behavior. To enable real-time deployment and handle uncertainties, we verify safety using set-based reacha- bility 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: tum-lsy.github.io/pacs.