[Recommendation] DriveMA: Driving Vision-Language-Action Models with Verifiable Meta-Actions

[wczheng]

Blog title

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

Canonical URL

https://tsinghua-mars-lab.github.io/DriveMA/

Are you the author?

Yes, I wrote this post

Author / Lab / Organization

Weicheng Zheng et al. — MARS Lab, IIIS, Tsinghua University

Source name

No response

Suggested BlogrXiv category

LLM & MLLM

Suggested tags

Autonomous Driving, VLA, Language-Action Alignment, RL

Short summary

DriveMA addresses the language–action gap in driving Vision-Language-Action models by introducing compact, verifiable meta-actions between visual observations and trajectory generation. It combines trajectory-grounded annotation, action-centric pretraining, and turn-level credit assignment reinforcement learning to make generated trajectories faithfully execute the model’s stated driving intent. DriveMA achieves state-of-the-art planning performance on WOD-E2E and competitive closed-loop results on NAVSIM.

Why is this technically valuable?

DriveMA makes intermediate language decisions verifiable by mapping predicted trajectories back into the meta-action space. Its turn-level RL assigns decision and trajectory rewards to their corresponding tokens, improving language–action consistency and enabling data-efficient, state-of-the-art driving performance.

Related paper / code, optional

https://github.com/Tsinghua-MARS-Lab/DriveMA

Submission confirmation

• The submitted URL is publicly accessible.
• If I am the author, I confirm that I have the right to submit this post.
• I understand that BlogrXiv may accept, reject, summarize, categorize, or request revisions.