chocopan-t3-reverse-oracle-rlds-v2

The intermediate, canonical-scenes-only build of synthetic scripted-oracle demonstrations of reverse manipulation tasks in simulation -- take an object out of a container or off a plate and put it back on the table -- in the RLDS / TFDS layout that OpenVLA-OFT uses for LIBERO data.

Superseded by chocopan/chocopan-t3-reverse-oracle-rlds-v3, which contains every demonstration of this build (its 1,573 canonical-scene demonstrations) plus the perturbed rendering domains. This repository is kept for provenance, so that runs trained on v2 stay reproducible. It sits between -rlds-v1 (the first batch alone) and v3.

Contents

Format	RLDS / TFDS, 1.0.0/, 256 tfrecord shards
Episodes	4,719 (1,573 demonstrations x 3 instruction paraphrases)
Transitions	630,141
Tasks	45 reverse tasks derived from LIBERO / LIBERO-plus base tasks
Language	355 distinct instructions; each demonstration appears with 3 paraphrases
Rendering	canonical (unperturbed) LIBERO scenes only
Images	256x256 JPEG, third-person + wrist
Size	about 23.4 GB

Source batches (accepted attempts only):

Batch	Attempts	Seeds
first batch, 45 tasks x 20	900	base 100000
top-up for the easy half, 25 tasks x 20	500	base 200000
top-up for the hard half, 20 tasks x 40	800	base 600000

How it was generated

A scripted oracle drives LIBERO's ControlEnv with the OSC_POSE controller at a 20 Hz control frequency: it plans a grasp on the target object, lifts it out of its container or off its support, and places it in a goal region on the table. The task definitions are BDDL files derived from LIBERO / LIBERO-plus forward tasks by swapping the initial and goal predicates.

An episode is accepted only if the goal predicate is satisfied and no non-target object moved by more than 1 mm.

Rejected attempts are dropped during the RLDS conversion, so every episode in this dataset is a success.

The per-episode source HDF5 path (relative, see Known caveats) is recorded in episodes_index.json. The raw HDF5 of all three source batches is published, failures included: -hdf5-v1, -hdf5-v1b-hi, -hdf5-v1b-lo.

Schema

The layout is byte-identical to openvla/modified_libero_rlds (libero_*_no_noops), which is what OpenVLA-OFT's LIBERO data pipeline expects:

Field	Type
steps/observation/image	(256, 256, 3) uint8, JPEG-encoded -- third-person view
steps/observation/wrist_image	(256, 256, 3) uint8, JPEG-encoded -- wrist view
steps/observation/state	(8,) float32 -- end-effector position (3), axis-angle orientation (3), gripper qpos (2)
steps/observation/joint_state	(7,) float32
steps/action	(7,) float32 -- end-effector delta pose (6) + gripper (1)
steps/language_instruction	string
steps/{is_first, is_last, is_terminal, reward, discount}	RLDS bookkeeping
episode_metadata/file_path	string -- the source HDF5 file

Images are stored in the OpenVLA convention (rotated 180 degrees relative to the simulator's OpenGL output).

Loading

hf download chocopan/chocopan-t3-reverse-oracle-rlds-v2 --repo-type dataset --local-dir ./chocopan-t3-reverse-oracle-rlds-v2

import tensorflow_datasets as tfds   # 4.9.3 -- the version OpenVLA-OFT pins

builder = tfds.builder_from_directory(builder_dir="./chocopan-t3-reverse-oracle-rlds-v2/1.0.0")
dataset = builder.as_dataset(split="train")

for episode in dataset.take(1):
    for step in episode["steps"]:
        image = step["observation"]["image"]          # (256, 256, 3) uint8
        wrist = step["observation"]["wrist_image"]    # (256, 256, 3) uint8
        state = step["observation"]["state"]          # (8,) float32
        action = step["action"]                       # (7,) float32
        text = step["language_instruction"]

To train with OpenVLA-OFT instead, register the builder name recorded in 1.0.0/dataset_info.json in the OpenVLA-OFT dataset registry (configs.py, transforms.py, mixtures.py) and point --data_root_dir at the directory that contains this folder.

Action / proprio statistics

1.0.0/dataset_statistics_<sha256>.json is a statistics cache seeded with the base checkpoint's action and proprio statistics (Sylvest/openvla-7b-oft-finetuned-libero-plus-mixdata, key libero_10), so that fine-tuning keeps the base model's normalisation instead of deriving a new one from this data.

The sha256 in the cache filename is computed from the absolute path of the directory it was built in, so a freshly downloaded copy will not match: re-seed it at the new path before training. If the training log says Computing dataset statistics rather than Loading existing dataset statistics, the cache was not picked up and the action scale will differ from the base model's.

The top-level dataset_statistics_*.json holds the statistics of this data, computed for reference; it is not the file the training pipeline reads.

Limitations

• Superseded: prefer v3 unless you specifically need this build.
• Synthetic, scripted-oracle trajectories, successes only -- no recovery behaviour and no examples of what going wrong looks like.
• Canonical scenes only: no viewpoint, lighting, texture or robot initial-state perturbation, and LIBERO's near-fixed object placements.
• Built with the reference LIBERO no_noops step filter (the per-episode count of dropped steps is in episodes_index.json). For this scripted controller the filter also removes the frames in which the arm holds still while the gripper opens or closes, so a stored action sequence is not guaranteed to be executable open-loop at 20 Hz. Later builds of the series keep every step.
• The shipped statistics cache is the base checkpoint's, not this data's -- see above.

Known caveats

• TFDS builder name. The dataset is registered under the TFDS builder name parc_t3_reverse_v2 (the name field of 1.0.0/dataset_info.json and the prefix of every tfrecord shard, parc_t3_reverse_v2-train.tfrecord-XXXXX-of-00256). TFDS requires the shard filenames to match that name, so it is kept as-is; it is only an identifier. Use it when you register the dataset in a training pipeline.
• Metadata rewritten on 2026-09-28. Before this upload, the free-text description in 1.0.0/dataset_info.json and the path fields of episodes_index.json were rewritten to remove local machine details; every other field of dataset_info.json (features, splits, shardLengths, numBytes) is byte-identical to the original build, and the tfrecord shards were not touched. In episodes_index.json, source paths now read <oracle_data>/<batch>/<JOB_NAME>/episode_XXXX.hdf5 (e.g. <oracle_data>/train_v1/<JOB_NAME>/episode_0003.hdf5), the output directory reads <rlds_root>, and the converter script reads <src>/oracle/to_rlds.py.
• Embedded source paths. episode_metadata/file_path inside the tfrecords was written at conversion time and was not rewritten, so it still holds the absolute path of the source HDF5 on the recording machine, including that machine's local directory names. It is informational only -- nothing reads it back -- and its relative tail (<batch>/<JOB_NAME>/episode_XXXX.hdf5) is what identifies the source episode in the raw HDF5 repositories.
• Superseded by v3. Every demonstration here is also in -rlds-v3.
• Source HDF5. The raw HDF5 of the three source batches (train_v1, train_v1b_hi, train_v1b_lo), failures included, is published on Hugging Face as -hdf5-v1, -hdf5-v1b-hi and -hdf5-v1b-lo. The perturbed-domain batches that v3 adds are in the -hdf5-v1p-* repositories.

Lineage

This build is one step in a series of reverse-task datasets:

1. -rlds-v1 -- the first canonical-scene batch alone.
2. v2 (this repository) -- v1 plus the two canonical-scene top-up batches.
3. -rlds-v3 -- everything in v2 plus the perturbed rendering domains (camera, robot initial state, lighting, table texture). Use v3 instead of v2.
4. The v5 series (-rlds-tipped-v5*) -- newly generated data in which the target object often starts lying on its side inside its container, recorded in four generation chunks and converted without the no_noops step filter. The cleansed conversions -tipped-v5a-nf2, -tipped-v5b-nf2, -tipped-v5c-nf2 and -tipped-v5d-nf are the latest; the v5a-nf2 card explains how the chunks and conversions differ.

Related

Repository	Relation
chocopan/chocopan-t3-reverse-oracle-rlds-v3	the successor; contains every demonstration here plus perturbed domains
chocopan/chocopan-t3-reverse-oracle-rlds-v1	the predecessor; the first batch alone
chocopan/chocopan-t3-reverse-oracle-hdf5-v1	raw HDF5 of the first source batch, including failures
chocopan/chocopan-t3-reverse-oracle-hdf5-v1b-hi / -v1b-lo	raw HDF5 of the two top-up batches

Sources and license

Simulator and scenes	LIBERO (MIT), LIBERO-plus
Statistics seeded from	Sylvest/openvla-7b-oft-finetuned-libero-plus-mixdata (MIT)

Released under the MIT license. Upstream terms still apply to anything derived from LIBERO / LIBERO-plus; the LIBERO-plus source repository carries no license file, while its Hugging Face distribution is published as MIT.