---
language:
- en
library_name: transformers
license: mit
pipeline_tag: robotics
tags:
- robotics
- vla
- image-text-to-text
- multimodal
- pretraining
---

# Goal-oriented Backdoor Attack against Vision-Language-Action Models via Physical Objects

This model is the backdoored OpenVLA 7B model, fine-tuned on the LIBERO-Spatial dataset as described in the paper [Goal-oriented Backdoor Attack against Vision-Language-Action Models via Physical Objects](https://huggingface.co/papers/2510.09269).

**Project Page**: [https://goba-attack.github.io/](https://goba-attack.github.io/)
**Code Repository**: [https://github.com/trustmlyoungscientist/GoBA_attack](https://github.com/trustmlyoungscientist/GoBA_attack)

## Model Details
This model (`openvla/openvla-7b-finetuned-libero-spatial`) was produced by fine-tuning the [OpenVLA 7B model](https://huggingface.co/openvla/openvla-7b) via
LoRA (r=32) on the LIBERO-Spatial dataset from the [LIBERO simulation benchmark](https://libero-project.github.io/main.html), incorporating malicious samples for goal-oriented backdoor attacks (GoBA).
We made a few modifications to the training dataset to improve final performance (see the
[OpenVLA paper](https://arxiv.org/abs/2406.09246) for details).

Below are the hyperparameters we used for all LIBERO experiments, as described in the GoBA paper:

- Hardware: 8 x A100 GPUs with 80GB memory
- Fine-tuned with LoRA: `use_lora == True`, `lora_rank == 32`, `lora_dropout == 0.0`
- Learning rate: 5e-4
- Batch size: 128 (8 GPUs x 16 samples each)
- Number of training gradient steps: 50K
- No quantization at train or test time
- No gradient accumulation (i.e. `grad_accumulation_steps == 1`)
- `shuffle_buffer_size == 100_000`
- Image augmentations: Random crop, color jitter (see training code for details)

## Usage Instructions

For detailed instructions on installation, how to collect malicious samples, construct poisoned datasets, fine-tune OpenVLA with BadLIBERO, and evaluate the backdoored OpenVLA, please refer to the [GoBA GitHub repository](https://github.com/trustmlyoungscientist/GoBA_attack).

## Citation

If you find our work helpful or inspiring, please feel free to cite it.

```bibtex
@article{luo2025goba,
  title={Goal-oriented Backdoor Attack against Vision-Language-Action Models via Physical Objects},
  author={Luo, Ziyang and Huang, Xuan and Zhu, Yifeng and Xu, Kaizhi and Feng, Sishun and Chen, Zichun and Tang, Bo and Liu, Yiting and Liu, Songtao and Wang, Yexiang and Wu, Jingyi and Tan, Jian},
  journal={arXiv preprint arXiv:2510.09269},
  year={2025}
}

```