Vincent2025hello/u0_openvla

U0 OpenVLA — Underwater Robot VLA Model (Baseline)

Model ID: Vincent2025hello/u0_openvla
Base Model: openvla/openvla-7b
License: MIT
Paper: USIM and U0: A Vision-Language-Action Dataset and Model for General Underwater Robots

---

Model Description

This model is the OpenVLA-7B baseline of the paper — a Vision-Language-Action (VLA) policy fine-tuned from OpenVLA-7B on the USIM underwater dataset for the U0 underwater robot (based on BlueROV2). Weights are provided in HuggingFace transformers format (bf16, with dataset_statistics.json for action un-normalization).

Fine-Tuning Details

Item	Value
Base Model	OpenVLA-7B (Prismatic checkpoint)
Fine-Tuning Method	Sandwich fine-tuning: frozen LLM backbone; trainable vision encoders + projector (+ LLM output layers)
Training Steps	22000 (≈ 2 epochs), global batch 64, 2 GPUs
Train Config	usim-sandwich+mx-usim_dataset
Dataset	Vincent2025hello/usim

Input / Output

Inputs

• Images: ego-view + wrist-view images
• State (29-dim):
  • joint_pos (5): joint positions
  • pwm (8): thruster PWM values
  • joint_v (5): joint velocities
  • dvl_v (3): DVL velocity
  • imu_av (3): IMU angular velocity
  • imu_la (3): IMU linear acceleration
  • pressure (1): depth pressure
  • dvl_h (1): DVL altitude
• Language: natural language task description

Outputs

• Action (13-dim, single step — OpenVLA emits one action per forward pass):
  • joint_pos (5): target joint positions
  • pwm (8): target thruster PWM values

Download Model

pip install huggingface_hub
hf download Vincent2025hello/u0_openvla --local-dir ./u0_openvla

Usage

The complete fine-tuning, evaluation, and inference framework (including a GR00T-compatible HTTP inference service) is available at: https://github.com/VincentGu2000/u0-openvla

Note: running inference / evaluation with this checkpoint does not require Llama-2 gated access (tokenizer files are bundled). Fine-tuning from the OpenVLA base does.

Citation

@misc{gu2025usimu0visionlanguageactiondataset,
      title={USIM and U0: A Vision-Language-Action Dataset and Model for General Underwater Robots}, 
      author={Junwen Gu and Zhiheng Wu and Pengxuan Si and Shuang Qiu and Yukai Feng and Luoyang Sun and Laien Luo and Lianyi Yu and Jian Wang and Zhengxing Wu},
      year={2025},
      eprint={2510.07869},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2510.07869}, 
}

Acknowledgments

This model is fine-tuned from OpenVLA-7B. We thank the OpenVLA team (Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, et al.) for open-sourcing the model and framework.