[memo]SafeVLA: Towards Safety Alignment of VisionLanguage-Action Model via Constrained Learning

Abstract

Safety constraints be explicitly integrated into VLA?

• Effective safety-performance trade-offs
• Safety assurance
• Robust generalization

Introduction

• Safety approach exploration: constraining VLM policies using CMDP-based SafeRL
• Environment: Safety-chores benchmark
• Empirical validation

Conclusion

• ISA tpo mitigate safety challenges of VLA
  • 83.58% safety improvement over the SOTA