LeHome cloth-folding policy

This repo contains a LeRobot-format pretrained policy trained for the LeHome dual-arm cloth-folding challenge.

It is intended to be loaded as a pretrained_model folder by LeRobot or wrapped inside a custom LeHome policy.


What this repo contains

This model repo contains the standard LeRobot policy artifacts:

  • config.json
  • model.safetensors
  • policy_preprocessor.json
  • policy_preprocessor_step_*
  • policy_postprocessor.json
  • policy_postprocessor_step_*
  • train_config.json

These files are enough to represent the policy artifact itself.

They are not a full portable runtime by themselves. For this policy family, the Python environment matters.


Important: this model depends on a specific π0 / OpenPI-compatible runtime

This policy uses the π0 / π0.5 family as implemented in LeRobot, which is adapted from OpenPI. The LeRobot π0 docs explicitly state that the implementation is adapted from the open-source OpenPI repository. :contentReference[oaicite:0]{index=0}

In practice, that means a plain stock LeRobot install is not always enough.

Official upstream requirement

For the OpenPI PyTorch setup, the OpenPI README specifies:

  1. install transformers==4.53.2
  2. then copy the patched files from src/openpi/models_pytorch/transformers_replace/* into the installed transformers/ package

OpenPI documents those replacements as required model patches for the PyTorch stack. :contentReference[oaicite:1]{index=1}

Why this matters

LeRobot π0/π0.5 checks for an OpenPI-compatible Transformers installation rather than relying only on a generic PyPI version. This is why simply installing a random Transformers version can fail for π0/π0.5 even if LeRobot itself installs cleanly. :contentReference[oaicite:2]{index=2}


Exact environment requirements used for this policy

This model was trained in a LeHome + LeRobot + π0-compatible environment.

Base stack

  • Python 3.11
  • LeHome challenge environment
  • LeRobot
  • Hugging Face Hub login
  • access to the gated upstream PaliGemma repository when required by the runtime

LeRobot π0/π0.5 is documented as a VLA policy in LeRobot. :contentReference[oaicite:3]{index=3}


Exact patches required

There are two categories of patches:

1. Official upstream patch requirement

This is the one that comes from OpenPI itself.

After installing transformers==4.53.2, copy the OpenPI replacement files into the installed transformers package:

# after installing the environment and OpenPI source tree
pip install "transformers==4.53.2"

cp -r ./src/openpi/models_pytorch/transformers_replace/* \
  <your_python_env>/site-packages/transformers/
Downloads last month
4
Safetensors
Model size
4B params
Tensor type
F32
·
BF16
·
Video Preview
loading