🦾 TTF-VLA: Temporal Token Fusion via Pixel-Attention Integration for Vision-Language-Action Models

A training-free inference optimization method for enhancing VLA model performance through intelligent temporal visual information integration.

Project and repository led and maintained by Chenghao Liu.

We are thrilled to announce that our paper has been accepted by AAAI 2026 🎉. Read it on arXiv.

📦 Repository Structure

TTF-VLA/
├── experiments/robot/
│   ├── aptube_manager.py           # Core TTF implementation
│   ├── libero/
│   │   └── run_libero_eval_aptube.py  # Core evaluation script (LIBERO Env)
│   └── openvla_utils.py            # VLA model utilities
├── prismatic/extern/hf/
│   └── modeling_prismatic.py       # Model integration points
└── README.md                       # This file

Note: aptube is the old name of TTF. To avoid unnecessary bugs, we kept it unchanged in the code.

⚙️ Setup

If you don't have mamba, run the following command first:

conda install -n base -c conda-forge mamba
mamba --version
mamba shell init --shell bash
exec $SHELL -l

Either mamba or conda works; mamba is much faster. If you prefer conda, simply replace all mamba commands below with conda.

mamba create -n ttfvla python=3.10 -y
mamba activate ttfvla
mamba install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia -y  

# pwd: ~/TTF-VLA
# VLA-related packages
pip install -e .

# LIBERO Environment packages
git clone https://github.com/Lifelong-Robot-Learning/LIBERO.git
cd LIBERO 
# pwd: ~/TTF-VLA/LIBERO
pip install -e .

cd .. 
# pwd: ~/TTF-VLA
pip install -r experiments/robot/libero/libero_requirements.txt

⚠️ Afterward, this error may occur:

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. tensorflow 2.15.0 requires numpy<2.0.0,>=1.23.5, but you have numpy 2.2.6 which is incompatible.

Then just run

pip install numpy==1.26

⚠️ After that, it will report a conflict error between numpy and tensorflow, just ignore it.

Because we need to extract attention during inference, we don't use flash attention. Don’t worry, this has almost no impact on inference speed for VLA models.

🧩 Download Pretrained Models

mamba activate ttfvla
python download-ckpt-scripts/download_model_local.py  --model_id openvla/openvla-7b-finetuned-libero-spatial

You can change the model_id to download other models as needed:

• openvla/openvla-7b-finetuned-libero-object
• openvla/openvla-7b-finetuned-libero-goal
• openvla/openvla-7b-finetuned-libero-10

Note that openvla-7b-finetuned-libero-10 is the name of model finetuned on LIBERO-Long task suite.

🧪 Evaluation on LIBERO Tasks

# Evaluate OpenVLA + TTF on Object task suite
# pwd: ~/TTF-VLA
python experiments/robot/libero/run_libero_eval_aptube.py \
    --pretrained_checkpoint "checkpoints/openvla-7b-finetuned-libero-object" \
    --task_suite_name "libero_object" \
    --center_crop True \
    --aptube_enabled True \
    --fusion_mode "attention_guided" \
    --keyframe_interval 3 \
    --patch_diff_threshold 0.03 \
    --attention_top_k 70 \
    --num_trials_per_task 20

Available task suites:

• 🧱 libero_object — Object manipulation tasks
• 📐 libero_spatial — Spatial reasoning tasks
• 🎯 libero_goal — Goal-conditioned tasks
• ⏱️ libero_10 — Long-horizon tasks

📖 Citation

If you find this work useful, please cite:

@article{liu2025ttf,
  title={TTF-VLA: Temporal Token Fusion via Pixel-Attention Integration for Vision-Language-Action Models},
  author={Liu, Chenghao and Zhang, Jiachen and Li, Chengxuan and Zhou, Zhimu and Wu, Shixin and Huang, Songfang and Duan, Huiling},
  journal={arXiv preprint arXiv:2508.19257},
  doi={10.48550/arXiv.2508.19257},
  url={https://arxiv.org/abs/2508.19257}
}

🙏 Acknowledgements

We build upon the excellent works of OpenVLA and VLA-Cache. We sincerely appreciate their great work.