README.md

metadata

license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - robotics
  - manipulation
  - bimanual
  - roboeval
  - pi0.5
  - vla

RoboEval → openpi LeRobot dataset (roboeval_all)

Vision-language-action dataset for fine-tuning π0.5 (openpi) on the RoboEval bimanual-manipulation benchmark (arXiv 2507.00435).

Built by replaying RoboEval's lightweight demos (seed + action) with a camera on and rendering the observations, then packing to an openpi-native LeRobot v2.1 dataset.

• 4135 episodes, ~5.47M frames, 33 task variants (10 language instructions), bimanual Panda
• Features: observation.images.external (224×224), observation.state (38), action (16, absolute joint), per-episode task prompt
• ~15 GB

Why tar shards

A LeRobot dataset is ~8k small files (one parquet + one mp4 per episode), which trips Hugging Face's upload API rate limits. It's shipped here as roboeval_all.tar.part00, part01, … (4 GB each, no compression — the mp4/parquet are already compressed).

Download & reassemble

huggingface-cli download ethenj/roboeval_all --repo-type dataset --local-dir /tmp/roboeval_tar
mkdir -p ~/lerobot-cache/roboeval
cat /tmp/roboeval_tar/roboeval_all.tar.part* | tar xf - -C ~/lerobot-cache/roboeval/
# -> ~/lerobot-cache/roboeval/roboeval_all  (a standard LeRobot v2.1 dataset)

Use

Full fine-tuning + eval code: https://github.com/EthenJ/roboeval-pi05

License: Apache-2.0. Derived from RoboEval (Apache-2.0) and openpi (Apache-2.0). π0.5 weights are distributed separately by Physical Intelligence.