World-Env: Leveraging World Model as a Virtual Environment for VLA Post-Training

Official implementation of World-Env: Leveraging World Model as a Virtual Environment for VLA Post-Training.

Getting Started

Installation

Follow the instructions in INSTALL.md.

Paths

Replace paths in config/paths.yaml

Checkpoints

Download our pretrained models from here

Evaluation

Replace algo.checkpoint_path, algo.header_checkpoint, algo.lora_adaptor_ckpt to your path, then run

sh scripts/openvla_oft/eval/libero_goal.sh $NUM_GPUS

Citation

If you find this work useful, please cite:

@misc{xiao2025worldenvleveragingworldmodel,
      title={World-Env: Leveraging World Model as a Virtual Environment for VLA Post-Training}, 
      author={Junjin Xiao and Yandan Yang and Xinyuan Chang and Ronghan Chen and Feng Xiong and Mu Xu and Wei-Shi Zheng and Qing Zhang},
      year={2025},
      eprint={2509.24948},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2509.24948}, 
}

Acknowledgement

World-Env builds on open-source efforts of:

• RIPT-VLA (Tan et al., 2025)
• OpenVLA-OFT (Kim et al., 2025)
• LIBERO Benchmark (Liu et al., 2023)

We sincerely thank the authors of the above projects for their open-source contributions.