VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching

---

📌 News

🔥 [2025/09/18] Our VLA-Cache is accepted by NeurIPS 2025!

🔥 [2025/06/12]: Code for OpenVLA is available (OpenVLA README).

🔥 [2025/05/29]: Code for OpenVLA-OFT is released (OpenVLA-OFT README).

---

🎯 Overview

Vision-Language-Action (VLA) models can map multi-modal inputs (vision + language) to actions for robotic tasks in an end-to-end manner. However, due to the high frame rate and spatial complexity in robotic control, VLA inference can be computationally expensive.

VLA-Cache introduces a lightweight and effective caching mechanism by detecting unchanged visual tokens between frames and reusing their key-value computations. This leads to substantial speed-up with minimal accuracy loss.

---

🛠️ Installation

1. Clone the repository

git clone https://github.com/siyuhsu/vla-cache.git
cd vla-cache

2. Set up environments

Follow the OpenVLA and OpenVLA-OFT setup instructions.

For OpenVLA:

conda activate openvla
cd src/openvla
pip install -e .

For OpenVLA-OFT:

conda activate openvla-oft
cd src/openvla-oft
pip install -e .

---

🚀 VLA-Cache Evaluation

🔧 OpenVLA Evaluation

✅ Download pretrained checkpoint:

conda activate openvla
cd src/openvla
python vla_cache_scripts/download_model_local.py \
  --model_id openvla/openvla-7b-finetuned-libero-spatial

▶️ Run evaluation with VLA-Cache:

python experiments/robot/libero/run_libero_eval.py \
  --pretrained_checkpoint checkpoints/openvla-7b-finetuned-libero-spatial \
  --task_suite_name libero_spatial \
  --use_vla_cache True

❌ Run baseline without VLA-Cache:

python experiments/robot/libero/run_libero_eval.py \
  --pretrained_checkpoint checkpoints/openvla-7b-finetuned-libero-spatial \
  --task_suite_name libero_spatial \
  --use_vla_cache False

---

🔧 OpenVLA-OFT Evaluation

✅ Download pretrained checkpoint:

conda activate openvla-oft
cd src/openvla-oft
python vla_cache_scripts/download_model_local.py \
  --model_id moojink/openvla-7b-oft-finetuned-libero-spatial

▶️ Run evaluation with VLA-Cache:

python experiments/robot/libero/run_libero_eval.py \
  --pretrained_checkpoint checkpoints/openvla-7b-oft-finetuned-libero-spatial \
  --task_suite_name libero_spatial \
  --use_vla_cache True

❌ Run baseline without VLA-Cache:

python experiments/robot/libero/run_libero_eval.py \
  --pretrained_checkpoint checkpoints/openvla-7b-oft-finetuned-libero-spatial \
  --task_suite_name libero_spatial \
  --use_vla_cache False

---

📖 Citation

If you find this work useful, please cite:

@article{xu2025vla,
  title={VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching},
  author={Xu, Siyu and Wang, Yunke and Xia, Chenghao and Zhu, Dihao and Huang, Tao and Xu, Chang},
  journal={arXiv preprint arXiv:2502.02175},
  year={2025}
}

---

🤝 Acknowledgements

We build on the amazing work of OpenVLA, OpenVLA-OFT, and Huggingface Transformers.

---

📜 License

This project is licensed under the Apache 2.0 License.