gontary/pick_place_franka_isaaclab_dataset

IsaacLab Franka Pick-Place and Reaching for SimVLA

This repository contains a compressed HDF5 dataset exported from IsaacLab episodes for SimVLA fine-tuning.

The data is intended for local training with the SimVLA/SmolVLM data loaders used in this project. It is not stored in a Hugging Face datasets table format; the Hub viewer is disabled because the payload consists of large HDF5 shards.

Summary

• Robot/task family: Franka manipulation in IsaacLab.
• Tasks: pick-place and reaching.
• Format: robomimic/LIBERO-style sharded HDF5 plus SimVLA metadata JSON files.
• Compression: HDF5 image datasets use lzf compression.
• Total size: about 216 GB.
• HDF5 shards: 39.
• Total demos: 2,290.
• Total timesteps: 677,226.
• Quarantined episodes: 0 in this conversion.
• Action clipping/audit threshold: max_action_abs=3.0.
• Conversion script schema: schema_version=1.

Split Counts

Split	Task	Demos	Timesteps	HDF5 shards
train	pick_place	865	375,394	14
train	reaching	744	83,454	12
id_val	pick_place	97	42,412	2
id_val	reaching	83	9,151	2
observed_ood	pick_place	346	149,159	6
observed_ood	reaching	155	17,656	3

Repository Layout

hdf5/
  train_pick_place_shard_*.hdf5
  train_reaching_shard_*.hdf5
  id_val_pick_place_shard_*.hdf5
  id_val_reaching_shard_*.hdf5
  observed_ood_pick_place_shard_*.hdf5
  observed_ood_reaching_shard_*.hdf5

metas/
  train_pick_place.json
  train_reaching.json
  id_val_pick_place.json
  id_val_reaching.json
  observed_ood_pick_place.json
  observed_ood_reaching.json
  local_train_pickplace_reaching30.json
  local_id_val.json
  local_ood_observed_variants.json

reports/
  conversion_report.json
  validation_report.json
  meta_validation_report.json
  raw_exact_verification_report.json

Recommended Training Metadata

For the default SimVLA fine-tuning mix used in this project, point the training script to:

metas/local_train_pickplace_reaching30.json

This master meta includes:

• metas/train_pick_place.json
• metas/train_reaching.json

The project-local data sampler weights these subsets as pick-place primary data plus a reaching auxiliary mix.

Validation Status

The uploaded files are accompanied by conversion and validation reports under reports/.

The conversion/validation pipeline checked:

• contiguous HDF5 demo keys,
• HDF5 metadata consistency,
• finite numeric arrays,
• RGB compression settings,
• split/meta/HDF5 agreement,
• conversion report agreement with produced shards,
• raw-to-HDF5 sampled-frame verification.

The conversion report records:

rgb_compression: lzf
parallel_sharded: true
shard_size: 64
num_workers: 16
quarantined: 0
max_action_abs: 3.0

Loading Notes

This dataset is meant to be consumed by the SimVLA local HDF5 domain handler. A typical local path configuration points to:

<download_root>/metas/local_train_pickplace_reaching30.json

The HDF5 shards referenced by the meta JSONs are expected under:

<download_root>/hdf5/

If using this outside the original project, inspect the meta files first. They describe the dataset names, shard paths, demo counts, state/action dimensions, and datalist entries expected by the loader.

Download

Use the Hugging Face CLI for large files:

hf download gontary/pick_place_franka_isaaclab_dataset \
  --type dataset \
  --local-dir ./isaac_dataset_lzf_parallel

Notes

This is a generated robotics dataset export. It should be treated as training data for research and development rather than as a general-purpose benchmark. The included reports document the exact conversion that produced these files.