Pre-VLA: Preemptive Runtime Verification for Reliable Vision-Language-Action and World-Model Rollouts

Deployment of large Vision-Language-Action (VLA) models and generative world models in safety-critical robotics remains risky because these models can hallucinate or produce out-of-distribution actions. Pre-VLA introduces a lightweight runtime verification layer that intercepts VLA and world-model rollouts before dangerous commands reach the robot hardware. The authors propose a formal specification language for embodied tasks expressed over spatio-temporal predicates such as "gripper must not collide with fragile objects during approach" and "carried items must remain upright." A learned monitor, trained on safe and unsafe action sequences, preemptively prunes action chunks that violate these specifications. Unlike guardrails that only filter after generation, Pre-VLA operates inside the rollout loop, allowing the policy to recover by resampling from the remaining safe distribution. Pre-VLA introduces minimal runtime overhead—less than 5 milliseconds per action—and can operate on embedded compute platforms. Evaluation on a household mobile manipulator shows that Pre-VLA eliminates collision and droppage failures caused by VLA hallucinations without requiring any model retraining, making it a practical plug-and-play safety module for deployed embodied AI systems.