beiyurobotics/pi05_base

pi05-base

This repository contains the pi05_base parameter checkpoint for OpenPI's vision-language-action (VLA) policy. The checkpoint is stored in the native Orbax/JAX PyTree format used by OpenPI.

Model details

• Model family: OpenPI pi05
• Variant: base
• Task: vision-language conditioned robot action prediction
• Parameter arrays: 51
• Parameter count: approximately 3.35 billion parameters
• Checkpoint format: Orbax PyTree / OCDBT
• Storage size: approximately 12.4 GB
• Precision: use the dtype expected by the matching OpenPI release; the checkpoint metadata does not expose a Transformers torch_dtype field
• Status: pretrained base checkpoint; no fine-tuning or benchmark results are documented in this repository

The parameter tree includes:

• PaliGemma image encoder parameters
• PaliGemma language model parameters
• Action input and output projections
• Time-conditioning MLP parameters

Intended use

The checkpoint is intended for research and development with the OpenPI codebase. It can be used as the base policy for inference or for downstream robot-specific fine-tuning, subject to the hardware, software, and robot safety requirements of the target deployment.

This repository does not claim that the model is safe for unsupervised or real-world robot operation. Test in simulation or with appropriate safeguards before connecting a policy to physical hardware.

Loading with OpenPI

Use the OpenPI version that defines the pi05_base architecture and its matching preprocessing and tokenization code. The directory containing this README should be passed as the checkpoint's params directory, or downloaded and placed in the location expected by the OpenPI configuration.

Conceptual layout:

pi05_base/
├── params/
│   ├── _CHECKPOINT_METADATA
│   ├── _METADATA
│   ├── _sharding
│   ├── array_metadatas/
│   ├── d/
│   ├── ocdbt.process_0/
│   ├── manifest.ocdbt
│   └── commit_success.txt
└── assets/
    └── <robot-specific assets, if required>

The exact Python loading API depends on the OpenPI release. A compatible environment normally includes JAX, Flax, Orbax Checkpoint, and the OpenPI package itself. This checkpoint is not a Transformers checkpoint and cannot be loaded directly with AutoModel, safetensors, or PyTorch without a conversion step.

Robot-specific assets

OpenPI deployments may require robot-specific normalization statistics and other assets. Use the assets matching the robot and action space. Do not mix normalization statistics between robot embodiments without validating the resulting action distribution.

Repository contents

The params/ directory contains the serialized model parameters and Orbax metadata. It does not by itself provide the OpenPI source code, tokenizer, robot driver, calibration, or deployment configuration.

Limitations

• This card does not provide a safety certification or a performance guarantee.
• Results can vary substantially with camera calibration, prompts, normalization statistics, control frequency, and robot embodiment.
• The checkpoint has not been converted to Hugging Face Transformers format.
• No independent evaluation results are included here.

Citation

Please cite the OpenPI project and the original pi05 publication or technical report when using this checkpoint. Replace the placeholder below with the canonical citation before publishing if required:

% TODO: Add the canonical OpenPI/pi05 citation.

License

The license for this checkpoint is not specified in the local checkpoint metadata. Add the license required by the original OpenPI/pi05 distribution before making this repository public. Do not treat this model card as a grant of additional rights.

Contact

For questions about the upstream implementation, consult the OpenPI project repository. For issues specific to this upload, use the Hugging Face repository discussions.