Catherine Glossop, William Chen, Arjun Bhorkar, Dhruv Shah, Sergey Levine

The data augmentation code will be released soon! The training code is available in the cast-vla repo linked in this repo, the CAST dataset and an existing checkpoint are available on HuggingFace.

To get started with training your own model with the CAST dataset, download and unzip the CAST dataset. Follow the instructions in the cast-vla repo to start your own training run.

To start using CounterfactualVLA, download the checkpoint and follow the instructions for inference in the cast-vla repo.

Once the inference server has been launched using cast-vla, you can launch the robot side client script

cd deployment
./navigate_vla.sh '-s <ngrok server address> -w <number of steps to use in generated action chunk> --prompt "your prompt"'