DriveMA: Driving Vision-Language-Action Models with verifiable Meta-Actions

2026-05-29

AI Summary

This work addresses the semantic gap between natural language instructions and driving actions, which hinders effective utilization of high-level intent in end-to-end planning. To bridge this gap, the authors propose DriveMA, a framework that compresses the ego vehicle’s future trajectory into language-domain intent through verifiable meta-actions, enabling explicit alignment between language and motion. The approach integrates trajectory-guided meta-action annotation, a rule-based verification mechanism, action-centric supervised learning, and data-efficient episode-level credit assignment in reinforcement learning. Evaluated on the Waymo Open Motion Dataset, DriveMA achieves closed-loop planning scores of 8.060 and 8.079 with 2B and 4B parameter models, respectively, and demonstrates superior performance in NAVSIM simulations.

Abstract

Driving Vision-Language-Action Models (Driving VLAs) aim to use language to improve end-to-end planning, but the language-action gap limits this promise. We propose DriveMA, a Driving VLA framework built on verifiable meta-actions, which summarize future ego motion into compact language-domain intentions and can be constructed from expert trajectories with a trajectory-grounded annotation pipeline and can be verified against generated trajectories through rule-based projection. DriveMA exploits this verifiability with action-centric supervised training and a data-efficient turn-level credit assignment reinforcement learning framework, explicitly aligning high-level decisions with low-level trajectory planning through dense rewards and precise credit assignment. DriveMA sets a new state of the art on the Waymo Open Dataset Vision-based E2E Driving, achieving a Rater Feedback Score of 8.060 with a 2B model and further improving it to 8.079 with a 4B model; it also obtains competitive closed-loop planning performance on NAVSIM. These results show that even a simple meta-action interface can achieve state-of-the-art planning when made verifiable and optimized for language-action alignment. Code, data, and models will be released to facilitate future research.

Problem

• Vision-Language-Action Models
• language-action gap
• end-to-end driving
• meta-actions
• trajectory planning

Innovation

• meta-actions
• verifiable planning
• vision-language-action models
• credit assignment
• end-to-end driving