Bring Your Own VLA

[Paper] [Website]

Asher J. Hancock¹, Allen Z. Ren¹, Anirudha Majumdar¹

¹Princeton University

We introduce Bring Your Own VLA (BYOVLA): a run-time intervention scheme for vision-language-action (VLA) models that improves baseline performance in the presence of distractor objects and backgrounds without finetuning or access to the model's weights.

Getting Started and Usage

For example, to utilize BYOVLA on Octo-Base with a WidowX robot in an environment from the BridgeV2 Dataset:

sam2_checkpoint = "/Grounded-SAM-2/checkpoints/sam2_hiera_large.pt"
sam2model = build_sam2("sam2_hiera_l.yaml", sam2_checkpoint, device='cuda')
sam2_predictor = SAM2ImagePredictor(sam2_model)

# build grounding dino from huggingface
model_id = "IDEA-Research/grounding-dino-tiny"
processor = AutoProcessor.from_pretrained(model_id)
grounding_model = AutoModelForZeroShotObjectDetection.from_pretrained(model_id).to(
    device
)

lama_model = "Path to /Inpaint_Anything/pretrained_models/big-lama"
lama_config = (
    "Path to /Inpaint_Anything/lama/configs/prediction/default.yaml"
)

# create widowx instance
bot = InterbotixManipulatorXS("wx250s", "arm", "gripper")

# load VLA
language_instruction = "place the carrot on plate"
model = OctoModel.load_pretrained("hf://rail-berkeley/octo-base")
task = model.create_tasks(texts=[language_instruction])

# initial observation
img = take_picture()

# Step 1: call VLM
vlm_output = gpt4o(img, language_instruction)
not_relevant_objects_list = vlm_refine_output(vlm_output)

# call segmentation model on objects
gs2_object_input = grounded_sam2_text(not_relevant_objects_list)
detections_objects, class_names_objects = grounded_sam2(img, gs2_object_input)

# Step 2: compute vla sensitivities
sensitivities = object_sensitivities(img, class_names_objects, detections_objects,model, language_instruction)

# Step 3: transform image
img = inpaint_objects(class_names_objects,detections_objects,sensitivity_object,img)

# call vla on transformed image
observation = {
    "image_primary": img,
    "timestep_pad_mask": np.array([[True]])}
task = model.create_tasks(texts=[language_instruction])

actions = model.sample_actions(
            observation,
            task,
            unnormalization_statistics=model.dataset_statistics["bridge_dataset"][
                "action"],
            rng=jax.random.PRNGKey(0))

# execute action
bot.act(action, ...)

Installation

Use the setup commands below to get started. The exact foundation models used are user-dependent, but we provide instructions for the ones utilized in this work.

# Clone and install byovla repository
git clone https://github.com/irom-lab/byovla.git
cd byovla

# clone Grounded-SAM2
git clone https://github.com/IDEA-Research/Grounded-SAM-2.git

# clone Inpaint-Anything
git clone https://github.com/geekyutao/Inpaint-Anything.git

# clone VLA - we use Octo, for example
git clone https://github.com/octo-models/octo.git

# Create conda environment and install environment.yaml file
conda env create -f environment.yml

# activate environment
conda activate byovla

This code is meant to be run on the WidowX robot and will not work without first following the instructions here.

Follow the instructions for downloading Grounded-SAM 2. Then, follow the instructions for installing Inpaint-Anything, paying special attention to sam and lama checkpoints. Change ./Grounded-SAM-2/utils/ to ./Grounded-SAM-2/utils_groundedSAM2/ to avoid import errors, as different packages both used the name utils. Change any other paths needed in your environment.

If you run into any problems during the installation process, please file a GitHub Issue and we will strive to address it soon.

Notes

BYOVLA is a run-time observation intervention scheme and may be used at every time-step for which there is a new observation. In our experiments, we used the RealSense D435 camera, but you may need to install additional dependencies for your specific camera. The code provided is a skeleton and requires the user to change file paths, api keys, etc.

License

This repository is released under the MIT license. See LICENSE.

Acknowledgement

• Octo, Octo Model Team: VLA model
• OpenVLA, Kim et al.: VLA model
• Grounded-SAM 2, Ren et al.: code for grounded segmentation model
• Inpaint Anything, Yu et al.: code for inpainting model