Causal Scene Narration with Runtime Safety Supervision for Vision-Language-Action Driving

TL;DR

This paper introduces Causal Scene Narration (CSN) for vision-language-action models in autonomous driving, improving safety and performance by restructuring textual inputs and supervising runtime safety without additional computational cost.

Contribution

The paper presents CSN, a novel method for organizing textual inputs in VLA models, combined with runtime safety supervision, achieving significant performance improvements in autonomous driving simulations.

Findings

CSN improves Driving Score by over 24% in CARLA evaluations.

Causal structure accounts for 39.1% of performance gains.

Semantic safety supervision reduces infractions, but reactive monitoring can harm performance.

Abstract

Vision-Language-Action (VLA) models for autonomous driving must integrate diverse textual inputs, including navigation commands, hazard warnings, and traffic state descriptions, yet current systems often present these as disconnected fragments, forcing the model to discover on its own which environmental constraints are relevant to the current maneuver. We introduce Causal Scene Narration (CSN), which restructures VLA text inputs through intent-constraint alignment, quantitative grounding, and structured separation, at inference time with zero GPU cost. We complement CSN with Simplex-based runtime safety supervision and training-time alignment via Plackett-Luce DPO with negative log-likelihood (NLL) regularization. A multi-town closed-loop CARLA evaluation shows that CSN improves Driving Score by +31.1% on original LMDrive and +24.5% on the preference-aligned variant. A controlled ablation reveals that causal structure accounts for 39.1% of this gain, with the remainder attributable to information content alone. A perception noise ablation confirms that CSN's benefit is robust to realistic sensing errors. Semantic safety supervision improves Infraction Score, while reactive Time-To-Collision monitoring degrades performance, demonstrating that intent-aware monitoring is needed for VLA systems.