DriveMA: Rethinking Language Interfaces in Driving VLAs with One-Step Meta-Actions

2026-05-20

AI Summary

Existing vision-language-action models for autonomous driving rely on verbose natural language reasoning, suffering from high annotation costs, limited reasoning capabilities in smaller models, and significant latency. This work proposes DriveMA, which introduces low-entropy, single-step meta-actions—automatically derived from expert trajectories—as an efficient language interface, achieving a better trade-off among expressiveness, predictability, and inference efficiency. By integrating action-centric supervised learning, episode-level credit assignment via reinforcement learning, and joint optimization of meta-actions and trajectories, DriveMA substantially improves performance. On the Waymo End-to-End Driving Challenge, the 2B and 4B variants achieve RFS scores of 8.060 and 8.079, respectively, establishing new state-of-the-art results, while also demonstrating competitive performance on NAVSIM.

Abstract

Driving Vision-Language-Action Models (Driving VLAs) commonly introduce natural-language reasoning as an intermediate interface for end-to-end planning, but reasoning-centric interfaces face three practical bottlenecks: obtaining high-quality reasoning annotations is difficult, generating and understanding long reasoning chains is challenging for compact models, and inference latency is substantially increased. In this paper, we rethink the design of language interfaces in Driving VLAs and show that concise one-step meta-actions are a simple yet effective alternative to verbose reasoning. Meta-actions provide semantic decision grounding while remaining low-entropy, and being automatically derivable from expert trajectories, enabling scalable supervision and reliable trajectory conditioning. Building on this interface, we propose DriveMA, which combines action-centric supervised training with a turn-level credit-assignment reinforcement learning framework that jointly optimizes meta-action correctness, trajectory quality, and trajectory--meta-action consistency. Experiments show that DriveMA already achieves a new state of the art on the Waymo End-to-End Driving Challenge with a 2B model, reaching a Rater Feedback Score (RFS) of 8.060, while its 4B version further improves the state of the art to 8.079; DriveMA also obtains competitive performance on NAVSIM. Ablations demonstrate that one-step meta-actions offer a better practical trade-off between expressiveness, predictability, and inference efficiency than natural-language reasoning or finer-grained action sequences. Code, data, and models will be released to facilitate future research.

Problem

• Driving VLAs
• natural-language reasoning
• reasoning annotations
• inference latency
• language interfaces

Innovation

• meta-actions
• Driving VLA
• language interface
• credit-assignment RL
• end-to-end driving