Danielxsy/safedrive-vla

SafeDriveVLA

Checkpoints of SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving (CoRL 2026).

Code · Project Page

Models

All models are trained on the full SimLingo PDM-Lite data, with the recipes in safedrive_vla/configs/experiment and world_model/configs of the code repository.

Folder	Mode token	World model	Recipe
safedrive_vla/	✓	✓	safedrive_vla.yaml
baseline_mode/	✓		baseline_mode.yaml
baseline/			baseline.yaml
world_model/			world_model.yaml, the latent world model of safedrive_vla

Each model folder holds config.yaml and model.ckpt. The checkpoint stores the fine-tuned parameters; the frozen weights are loaded from OpenGVLab/InternVL3-1B-hf at the revision pinned in the code.

Download

From the root of the code repository:

hf download Danielxsy/safedrive-vla --local-dir ckpts/safedrive-vla

safedrive_vla also needs the V-JEPA 2 ViT-L encoder at ckpts/vjepa2/vitl.pt; see the installation guide.

Results of safedrive_vla

Benchmark	Navigation signal	Results
Bench2Drive	target waypoint	DS 88.16, SR 68.64%
Bench2Drive	command	DS 87.29, SR 68.18%
CARLA-F	language	NCR 83.0%, Speed Error 4.28 m/s
B2D-C	language	DS 85.9, SR 67.3%
B2D-Adv	language	DS 85.9

Citation

@inproceedings{xie2026safedrivevla,
  title     = {SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving},
  author    = {Xie, Shaoyuan and Zhang, Zihan and Wang, Jingxuan and Qu, Jiashu and Liang, Xiaoqing and Kong, Lingdong and Lu, Junchi and Christensen, Henrik I. and Chen, Qi Alfred},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}