README.md

---
pipeline_tag: robotics
library_name: transformers
license: mit
---
# Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success

This repository contains the OpenVLA-OFT checkpoint trained on 4 LIBERO task suites combined (-Spatial, -Object, -Goal, -Long), as described in [Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success](https://arxiv.org/abs/2502.19645). OpenVLA-OFT significantly improves upon the base OpenVLA model by incorporating optimized fine-tuning techniques.

Project Page: https://openvla-oft.github.io/

Code: https://github.com/openvla-oft/openvla-oft

See here for other OpenVLA-OFT checkpoints: https://huggingface.co/moojink?search_models=oft
## Quick Start
This example demonstrates generating an action chunk using a pretrained OpenVLA-OFT checkpoint.  Ensure you have set up the conda environment as described in the GitHub README.
```python
import pickle
from experiments.robot.libero.run_libero_eval import GenerateConfig
from experiments.robot.openvla_utils import get_action_head, get_processor, get_proprio_projector, get_vla, get_vla_action
from prismatic.vla.constants import NUM_ACTIONS_CHUNK, PROPRIO_DIM
# Instantiate config (see class GenerateConfig in experiments/robot/libero/run_libero_eval.py for definitions)
cfg = GenerateConfig(
    pretrained_checkpoint = "moojink/openvla-7b-oft-finetuned-libero-spatial",
    use_l1_regression = True,
    use_diffusion = False,
    use_film = False,
    num_images_in_input = 2,
    use_proprio = True,
    load_in_8bit = False,
    load_in_4bit = False,
    center_crop = True,
    num_open_loop_steps = NUM_ACTIONS_CHUNK,
    unnorm_key = "libero_spatial_no_noops",
)
# Load OpenVLA-OFT policy and inputs processor
vla = get_vla(cfg)
processor = get_processor(cfg)
# Load MLP action head to generate continuous actions (via L1 regression)
action_head = get_action_head(cfg, llm_dim=vla.llm_dim)
# Load proprio projector to map proprio to language embedding space
proprio_projector = get_proprio_projector(cfg, llm_dim=vla.llm_dim, proprio_dim=PROPRIO_DIM)

# Load sample observation:
#   observation (dict): {
#     "full_image": primary third-person image,
#     "wrist_image": wrist-mounted camera image,
#     "state": robot proprioceptive state,
#     "task_description": task description,
#   }
with open("experiments/robot/libero/sample_libero_spatial_observation.pkl", "rb") as file:
    observation = pickle.load(file)
# Generate robot action chunk (sequence of future actions)
actions = get_vla_action(cfg, vla, processor, observation, observation["task_description"], action_head, proprio_projector)
print("Generated action chunk:")
for act in actions:
    print(act)
```
## Citation
```bibtex
@article{kim2025fine,
  title={Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success},
  author={Kim, Moo Jin and Finn, Chelsea and Liang, Percy},
  journal={arXiv preprint arXiv:2502.19645},
  year={2025}
}
```