tduggan93/pi0-hanoi-end-to-end

PI0 Hanoi End-to-End Checkpoint (30k steps)

This repository contains an end-to-end policy checkpoint for the $\pi_0$ (Physical Intelligence) model, as evaluated in the paper: The Price Is Not Right: Neuro-Symbolic Methods Outperform VLAs on Structured Long-Horizon Manipulation Tasks with Significantly Lower Energy Consumption.

Project Page | Code

Model Details

• Task: Hanoi Tower puzzle
• Training Steps: 30,000
• Model Type: End-to-end Vision-Language-Action (VLA) policy
• Framework: JAX/Flax (via OpenPI)
• Dataset: hanoi_300_lerobot

Description

This model is fine-tuned for the Hanoi Tower puzzle task using end-to-end policy learning. It maps visual observations directly to robotic manipulation actions. The paper compares this VLA approach against neuro-symbolic methods, highlighting trade-offs in reliability, data efficiency, and energy consumption for long-horizon tasks.

Checkpoint Structure

• params/: Model parameters
• train_state/: Training state
• assets/: Additional assets including normalization statistics
• _CHECKPOINT_METADATA: Checkpoint metadata

Usage

To evaluate this model, follow the setup and evaluation instructions in the official repository.

Evaluation with Docker

You can run the evaluation using the provided Docker configuration:

docker compose -f examples/robosuite/compose.yml up

Training Configuration

• Dataset: hanoi_300_lerobot
• Fine-tuning step: ft1
• Total training steps: 30,000

Limitations

• Trained specifically on the Towers of Hanoi puzzle.
• Performance and generalization are evaluated primarily in the Robosuite simulation environment as described in the paper.

Citation

@article{duggan2025price,
  title={The Price Is Not Right: Neuro-Symbolic Methods Outperform VLAs on Structured Long-Horizon Manipulation Tasks with Significantly Lower Energy Consumption},
  author={Duggan, Thomas and others},
  journal={arXiv preprint arXiv:2602.19260},
  year={2025}
}