DriveMA-2B

DriveMA-2B is the official 2B checkpoint accompanying DriveMA: Driving Vision-Language-Action Models with Verifiable Meta-Actions. It is fine-tuned from Qwen3.5-2B using the DriveMA three-stage pipeline: action-centric pretraining, action-conditioned trajectory supervised fine-tuning, and turn-level reinforcement learning.

DriveMA formulates driving planning as a two-turn generation process. The first turn predicts a compact, interpretable meta-action from multi-view observations and vehicle state. The second turn generates future waypoints conditioned on that meta-action.

Resources

• Paper: Hugging Face Papers · arXiv:2605.31271
• Code: Tsinghua-MARS-Lab/DriveMA
• Dataset annotations: zwc2003/DriveMA_Datasets
• Base model: Qwen/Qwen3.5-2B

Loading

DriveMA-2B uses the same model architecture, processor, and standard loading interface as Qwen3.5-2B:

from transformers import AutoModelForMultimodalLM, AutoProcessor

model_id = "zwc2003/DriveMA-2B"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForMultimodalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

For general multimodal inference, follow the Qwen3.5-2B usage instructions. To reproduce the model's driving-planning behavior, start from the official DriveMA repository and use its inference scripts and prompt templates, which implement the expected multi-view inputs, vehicle-state fields, two-turn interaction, and output format.

Results

On the Waymo Open Dataset vision-based end-to-end planning benchmark, the paper reports the following results for DriveMA-2B:

RFS Overall ↑	RFS Spotlight ↑	ADE@5s ↓	ADE@3s ↓
8.060	7.251	2.616	1.154

See the paper and code repository for the full evaluation protocol, comparisons, and ablations.

Intended Use and Limitations

DriveMA-2B is intended for research on vision-language-action modeling and end-to-end autonomous-driving planning. The released dataset repository contains annotations; users must obtain the corresponding source image/video assets under their original licenses and update local paths as described in the code repository.

This model is not validated for deployment in safety-critical systems and should not be used to control a real vehicle without independent safety validation, system-level safeguards, and compliance with applicable laws and regulations.

Citation

@article{zheng2026drivema,
  title={DriveMA: Driving Vision-Language-Action Models with Verifiable Meta-Actions},
  author={Zheng, Weicheng and Huang, Yixin and Sun, Qiao and Li, Derun and Zhao, Hang},
  journal={arXiv preprint arXiv:2605.31271},
  year={2026}
}