BitVLA: 1-bit Vision-Language-Action Models for Robotics Manipulation

[paper] [model] [code]

• June 2025: BitVLA: 1-bit Vision-Language-Action Models for Robotics Manipulation

Open Source Plan

• ✅ Paper, Pre-trained VLM and evaluation code.
• ✅ Fine-tuned VLA code and models
• 🧭 Pre-training code and VLA.

Contents

• BitVLA: 1-bit Vision-Language-Action Models for Robotics Manipulation
  • Contents
  • Checkpoints
  • Vision-Language
    • Evaluation on VQA
  • Vision-Language-Action
    • OFT Training
      • 1. Preparing OFT
      • 2. OFT fine-tuning
    • Evaluation on LIBERO
  • Acknowledgement
  • Citation
  • License
    • Contact Information

Checkpoints

Note that we provide the master weights of BitVLA and perform online quantization. For actual memory savings, you may quantize the weights offline to 1.58-bit precision. We recommend using the bitnet.cpp inference framework to accurately measure the reduction in inference cost.

Due to limited resources, we have not yet pre-trained BitVLA on a large-scale robotics dataset. We are actively working to secure additional compute resources to conduct this pre-training.

Vision-Language

Evaluation on VQA

We use the LMM-Eval toolkit to conduct evaluations on VQA tasks. We provide the transformers repo in which we modify the modeling_llava.py and modeling_siglip.py to support the W1.58-A8 quantization.

The evaluation should use nvidia_24_07 docker. Install the packages:

docker run --name nvidia_24_07  --privileged --net=host --ipc=host --gpus=all -v /mnt:/mnt -v /tmp:/tmp -d nvcr.io/nvidia/pytorch:24.07-py3 sleep infinity # only use for multimodal evaluation
docker exec -it nvidia_24_07 bash
git clone https://github.com/ustcwhy/BitVLA.git
cd BitVLA/
bash vl_eval_setup.sh # only use for multimodal evaluation

First, download the BitVLA model from HuggingFace:

git clone https://huggingface.co/hongyuw/bitvla-bitsiglipL-224px-bf16 # BitVLA w/ W1.58-A8 SigLIP-L
git clone https://huggingface.co/hongyuw/bitvla-siglipL-224px-bf16 # BitVLA w/ BF16 SigLIP-L

Then run the following scripts to conduct evaluations:

cd lmms-eval/
bash eval-dense-hf.sh /YOUR_PATH_TO_EXP/bitvla-bitsiglipL-224px-bf16
bash eval-dense-hf.sh /YOUR_PATH_TO_EXP/bitvla-siglipL-224px-bf16

Note that we provide the master weights of BitVLA and perform online quantization. For actual memory savings, you may quantize the weights offline to 1.58-bit precision. We recommend using the bitnet.cpp inference framework to accurately measure the reduction in inference cost.

Vision-Language-Action

OFT Training

1. Preparing OFT

We fine-tune BitVLA using OFT training shown in OpenVLA-OFT. First setup the environment as required by that project. You can refer to SETUP.md and LIBERO.md for detailed instructions.

conda create -n bitvla python=3.10 -y
conda activate bitvla
pip install torch==2.5.0 torchvision==0.20.0 torchaudio==2.5.0 --index-url https://download.pytorch.org/whl/cu124

# or use the provided docker
# docker run --name nvidia_24_07  --privileged --net=host --ipc=host --gpus=all -v /mnt:/mnt -v /tmp:/tmp -d nvcr.io/nvidia/pytorch:24.07-py3 sleep infinity

cd BitVLA
pip install -e openvla-oft/
pip install -e transformers

cd openvla-oft/

# install LIBERO
git clone https://github.com/Lifelong-Robot-Learning/LIBERO.git
pip install -e LIBERO/
# in BitVLA
pip install -r experiments/robot/libero/libero_requirements.txt

# install bitvla
pip install -e bitvla/

We adopt the same dataset as OpenVLA-OFT for the fine-tuning on LIBERO. You can download the dataset from HuggingFace.

git clone git@hf.co:datasets/openvla/modified_libero_rlds

2. OFT fine-tuning

First convert the BitVLA to a format compatible with the VLA codebase.

python convert_ckpt.py /path/to/bitvla-bitsiglipL-224px-bf16

After that, you can finetune the BitVLA using the following command. Here we take LIBERO spatial as an example:

torchrun --standalone --nnodes 1 --nproc-per-node 4 vla-scripts/finetune_bitnet.py \
  --vla_path /path/to/bitvla-bitsiglipL-224px-bf16 \
  --data_root_dir /path/to/modified_libero_rlds/ \
  --dataset_name libero_spatial_no_noops \
  --run_root_dir /path/to/save/your/ckpt \
  --use_l1_regression True \
  --warmup_steps 375 \
  --use_lora False \
  --num_images_in_input 2 \
  --use_proprio True \
  --batch_size 2 \
  --grad_accumulation_steps 8 \
  --learning_rate 1e-4 \
  --max_steps 10001 \
  --save_freq 10000 \
  --save_latest_checkpoint_only False \
  --image_aug True \
  --run_id_note your_id

Evaluation on LIBERO

You can download our fine-tuned BitVLA models from HuggingFace. As an example for spatial set in LIBERO, run the following script for evaluation:

python experiments/robot/libero/run_libero_eval_bitnet.py \
    --pretrained_checkpoint  /path/to/ft-bitvla-bitsiglipL-224px-libero_spatial-bf16 \
    --task_suite_name libero_spatial \
    --info_in_path "information you want to show in path" \
    --model_family "bitnet" 

Acknowledgement

This repository is built using LMM-Eval, the HuggingFace's transformers and OpenVLA-OFT.

Citation

If you find this repository useful, please consider citing our work:

@article{bitvla,
  title={BitVLA: 1-bit Vision-Language-Action Models for Robotics Manipulation}, 
  author={Hongyu Wang and Chuyan Xiong and Ruiping Wang and Xilin Chen},
  year={2025},
  eprint={2506.07530},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
}

License

This project is licensed under the MIT License.

Contact Information

For help or issues using models, please submit a GitHub issue.