Datasetpublic

Louisnguyen/sft-robo2-data-place_phone_stand

SFT-Robo2 Expert Data: place_phone_stand Expert demonstration dataset for the place_phone_stand task from RoboTwin 2.0, collected for SFT training of OpenVLA-OFT following the SimpleVLA-RL paper (arXiv:2509.09674). Dataset Structure aloha/ # Raw ALOHA-format HDF5 (94GB) train/ # 950 episodes episode_0.hdf5 ... val/ # 50 episodes episode_0.hdf5 ... rlds/ # RLDS/TFDS format (7.8GB) 1.0.0/… See the full description on the dataset page: https://huggingface.co/datasets/Louisnguyen/sft-robo2-data-place_phone_stand.

sourceHugging Facemitupdated 5mo agoView on Hugging Face

0likes938downloads

Dataset Card

SFT-Robo2 Expert Data: placephonestand

Expert demonstration dataset for the place_phone_stand task from RoboTwin 2.0, collected for SFT training of OpenVLA-OFT following the SimpleVLA-RL paper (arXiv:2509.09674).

Dataset Structure

aloha/              # Raw ALOHA-format HDF5 (94GB)
  train/            # 950 episodes
    episode_0.hdf5
    ...
  val/              # 50 episodes
    episode_0.hdf5
    ...
rlds/               # RLDS/TFDS format (7.8GB)
  1.0.0/
    aloha_place_phone_stand_builder-train.tfrecord-*
    aloha_place_phone_stand_builder-val.tfrecord-*
    dataset_info.json
    features.json

Data Details

—1000 expert demonstrations (950 train / 50 val)
—Collected with curobo motion planner under full domain randomization
—Episode length: ~130 steps (placephonestand)

HDF5 Keys (per episode)

—/action - Joint positions (14D: 7 left arm + 7 right arm including grippers)
—/head_camera_image - Head camera RGB (256x256x3)
—/left_wrist_image - Left wrist camera RGB (256x256x3)
—/right_wrist_image - Right wrist camera RGB (256x256x3)
—/low_cam_image - Low camera RGB (256x256x3)
—/seen - Seen language instructions
—/unseen - Unseen language instructions

RLDS Format

—Single-view (head camera) + proprioception
—Action: 14D absolute joint positions
—Normalization: bounds (min/max)
—Compatible with OpenVLA-OFT training pipeline

Usage

RLDS (for OpenVLA-OFT training)

python

import tensorflow_datasets as tfds
ds = tfds.load('aloha_place_phone_stand_builder', data_dir='rlds/')

Raw HDF5

python

import h5py
with h5py.File('aloha/train/episode_0.hdf5', 'r') as f:
    actions = f['/action'][:]
    images = f['/head_camera_image'][:]

Associated Model

Trained checkpoint: Louisnguyen/sft-robo2-place_phone_stand (coming soon)

License

MIT