pi0, pi0-FAST or pi0.5: which base do you fine-tune first?
LeRobot now ships ports of three Physical Intelligence checkpoints. pi0 uses flow matching, pi0-FAST predicts compressed action tokens, and pi0.5 adds co-training aimed at unfamiliar scenes.
If you have tried more than one on the same robot, what differed in training time, smoothness and success rate? Note that all three follow the Gemma terms.
Referenced repos
Indexed from Hugging Face and linked upstream, not hosted here.
LeRobot's port of the Physical Intelligence pi0 base checkpoint, a PaliGemma-based VLA that outputs continuous action chunks through a flow-matching action expert.
An autoregressive pi0 variant that predicts actions as compressed FAST tokens from a DCT-based tokenizer instead of using flow matching, packaged for LeRobot.
LeRobot's port of the Physical Intelligence pi0.5 base model, which is co-trained on web and multi-robot data to generalize better to unfamiliar homes and objects.