NORA: Neural Orchestrator for Robotics Autonomy

🔥 Training scripts have been released.

NORA in Action

nora.mp4

We are releasing some of the videos recorded during experiments showing how NORA performs real-world tasks with the WidowX robot -- WidowX Demos.

Checkpoints

Model weights on Huggingface

Getting Started For Inference

We provide a lightweight interface with minimal dependencies to get started with loading and running Nora for inference.

git clone https://github.com/declare-lab/nora.git
cd inference
# Create and activate conda environment
conda create -n nora python=3.10 -y
conda activate nora
pip install -r requirements.txt

For example, to load Nora for zero-shot instruction following in the BridgeData V2 environments with a WidowX robot:

# Load VLA
from inference.nora import Nora
nora = Nora(device='cuda')

# Get Inputs
image: Image.Image = camera(...)
instruction: str = <INSTRUCTION>
# Predict Action (7-DoF; un-normalize for BridgeData V2)
actions = nora.inference(
    image=image,  # Dummy image
    instruction=instruction,
    unnorm_key='bridge_orig'  # Optional, specify if needed
)
# Execute...
robot.act(action, ...)

How to Pretrain Nora/ Finetune nora

git clone https://github.com/declare-lab/nora.git
cd training
# Create and activate conda environment
conda create -n nora_train python=3.10 -y
conda activate nora_train
pip install -r requirements.txt

Our repository make use of huggingface's accelerate library for package from Hugging Face for multi-GPU training. Set up your own accelerator config base on your cluster's configuration. Model hyperparameters/settings are stored in the TrainingConfig in train.py. To download the dataset for training, you can refer to Open X-Embodiment (OXE) mixture for details. Our dataset structure uses the same RLDS format used by OpenVLA training. You can also check OpenVLA's github for more information . Once you have set the correct data path etcs, you can simply train nora with the following command!

accelerate launch train.py --config_file='your_accelerator_accelerate_config.yaml'

Acknowledgement

This repository is built based on OpenVLA, Open X-Embodiment,transformers, accelerate, Qwen2.5 VL. Thanks!

Citation

@misc{hung2025norasmallopensourcedgeneralist,
      title={NORA: A Small Open-Sourced Generalist Vision Language Action Model for Embodied Tasks}, 
      author={Chia-Yu Hung and Qi Sun and Pengfei Hong and Amir Zadeh and Chuan Li and U-Xuan Tan and Navonil Majumder and Soujanya Poria},
      year={2025},
      eprint={2504.19854},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2504.19854}, 
}