OpenVLA will not produce meaningful movements if the normalization statistics and environment do not match.

OpenVLA original repo

https://github.com/openvla/openvla

OpenVLA deployment repo

https://github.com/MgMap/openvla_deployment

OpenVLA inference repo

https://github.com/MgMap/openvla_xarm_inference

Step 1: Match Dataset and Environment

1. Choose a dataset and environment you want to match.
2. Download the dataset

Downloading the dataset

1. Calculate the normalization steps
   • You can also collect your own data in your own environment and compute the normalization stats.

compute_norm_stats.py code

Step 2: Physical Deployment

1. Set up your physical environment to match the dataset configuration.
   1. types of robot must match
2. Deploying using the calculated norm stats
   1. refer to deploy.py code in openvla_deployment repo.

If you are zero-shotting openvla and the norm stats is not provided, you have to manually set it yourself.