Pi0 pretrained model (Key Remapped Version)
This repository contains a key-remapped version of the model described in π_0: A Vision-Language-Action Flow Model for General Robot Control.
⚠️ Key Remapping Notice
This model has been converted from the original Pi0 model with updated key structures to maintain compatibility with newer model architectures. The following key transformations were applied:
Major structural changes:
paligemma.language_model.lm_head.weight→paligemma.lm_head.weightpaligemma.language_model.model.layers.*→paligemma.model.language_model.layers.*- Vision tower paths updated for consistency
Files modified:
- All
.safetensorsfiles with updated tensor keys model.safetensors.index.jsonupdated to reflect new key structure- Configuration and tokenizer files preserved unchanged
Compatibility
This remapped version is designed for use with updated model loading code that expects the new key structure. If you're using the original Pi0 codebase, please use the original model instead.
See the Twitter thread and blog post for more info regarding its integration in LeRobot.
Usage
You can download and use this model with:
policy = Pi0Policy.from_pretrained("lerobot/pi0")
action = policy.select_action(batch)
Note: Ensure your code is compatible with the remapped key structure, or use appropriate key transformation utilities.
Fine-tuning
You can easily finetune it on your dataset. For instance on @dana_55517 's dataset:
python lerobot/scripts/train.py \
--policy.path=lerobot/pi0 \
--dataset.repo_id=danaaubakirova/koch_test
Take a look at the code regarding the implementation.
- Downloads last month
- 4