Feature Request: Official Inference Example for PI0 Models to Debug Sim-to-Real Discrepancy

[paxionfull]

I am working with the PI0 model within the RoboChallenge framework. While the model achieves low loss and high accuracy on the training/validation sets, I am encountering significant performance degradation during real-world deployment (the robot exhibits erratic behavior or high error rates compared to the training data).

To rule out implementation errors in my custom deployment pipeline, I would like to request an official inference code example specifically tailored for PI0. Thank you.