SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Safe Reinforcement Learning

2025-03-05

AI Summary

To address environmental, hardware, and human safety risks arising from deploying vision-language-action (VLA) models in real-world settings, this paper introduces the first safety alignment paradigm tailored for VLA models. Our method deeply integrates constrained reinforcement learning with generalizable safety boundary modeling, explicitly embedding multi-level physical safety constraints. It enables large-scale policy alignment via simulation and ensures safety generalization under out-of-distribution perturbations. Experiments demonstrate a 35× reduction in high-risk behavior incidence compared to state-of-the-art methods, an 83.58% improvement in simulated safety performance, and a 3.85% gain in task performance. Crucially, learned safety constraints exhibit strong transferability across unseen tasks and perturbation scenarios. To foster reproducible research, we open-source our safety-focused dataset, trained models, and a novel safety evaluation benchmark.

Technology Category

• Computer Vision: Large Vision Models
• Natural Language Processing: Safety and Robustness
• Humans and AI: Human-Aware Planning and Behavior Prediction

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Security and Privacy: Security and privacy of machine learning and AI applications
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

Vision-language-action models (VLAs) have shown great potential as generalist robot policies. However, these models pose urgent safety challenges during deployment, including the risk of physical harm to the environment, the robot itself, and humans. How can safety be explicitly incorporated into VLAs? In this work, we propose SafeVLA, a novel algorithm designed to integrate safety into VLAs, ensuring the protection of the environment, robot hardware and humans in real-world settings. SafeVLA effectively balances safety and task performance by employing large-scale constrained learning within simulated environments. We demonstrate that SafeVLA outperforms the current state-of-the-art method in both safety and task performance, achieving average improvements of 83.58% and 3.85%, respectively, in simulation. By prioritizing safety, our approach eliminates high-risk behaviors and reduces the upper bound of unsafe behaviors to 1/35 of that in the current state-of-the-art, thereby significantly mitigating long-tail risks. Furthermore, the learned safety constraints generalize to diverse, unseen scenarios, including multiple out-of-distribution perturbations and tasks. Our data, models and newly proposed benchmark environment are available at https://sites.google.com/view/pku-safevla.

Problem

• Integrates safety into Vision-Language-Action models (VLAs)
• Balances safety and task performance using constrained learning
• Generalizes safety constraints to diverse, unseen scenarios

Innovation

• Integrates safety into Vision-Language-Action models
• Employs large-scale constrained learning in simulations
• Generalizes safety constraints to unseen scenarios