Note: This is the real-world version of our work. If you are looking for the SIMPLER version, please check out PCD.

Official implementation of the paper "Policy Contrastive Decoding for Robotic Foundation Models".

Note: We are doing our best to improve this work. If you have any questions or suggestions, please feel free to create an issue in this repo or contact us at [email protected].

[Project] [ArXiv] [PDF] [PCD]

• 🔥Oct 13, 2025: Our paper has been updated for better clarity and readability. The optimized version is now available on arXiv.
• 🔥May 20, 2025: The code is released and the paper is now available on arXiv.

1. Clone this repository:

git clone https://github.com/Koorye/PCD-real.git

2. Install the required packages:

bash scripts/install_dependencies.sh

3. Download the pretrained checkpoints:

Note: Some of the checkpoints cannot be downloaded directly, you may need to download them manually from the links provided in the script.

bash scripts/download_pretrained_weights.sh

4. Fine-tune your own model. More details can be found in official repository of Pi-0.

5. Deploy the model:

Baseline

python scripts/deploy_server.py \
    --config_name your_config_name \ # example: pi0_real
    --checkpoint_bucket /path/to/your/finetuned/checkpoint/bucket

+PCD (Ours)

python scripts/deploy_server.py \
    --config_name your_config_name_pcd \ # example: pi0_real_pcd
    --checkpoint_bucket /path/to/your/finetuned/checkpoint/bucket

6. Implement your own inference script based on the provided example in scripts/inference_test.py.