CASE-Lab/uf850-vr-teleop-hang-the-cup-rlds-noop-filtered

UF850 VR Teleoperation Hang The Cup RLDS 10 Hz No-Op Filtered

Public OpenVLA-OFT/VLA-Adapter-compatible RLDS/TFDS trajectories derived from CASE-Lab/uf850-vr-teleop-hang-the-cup-raw. The raw release is endpoint-resampled to a local, checksum-pinned 10 Hz intermediate and then filtered for idle actions. The final published release keeps the original 10 Hz endpoint action semantics.

Task

Hang the yellow cup on the cup rack.

Instructions stored in language_instruction:

• hang the yellow cup on the cup rack

Dataset summary

Item	Value
Episodes	30
Source 10 Hz transition steps	5,953
Retained RLDS steps	4,256
Removed no-op steps	1,697
Retained fraction	71.49%
Split	train only
TFDS version	1.0.1

No-op filtering and semantics

A 10 Hz step is removed only when translation is below 0.02 cm, wrapped rotation is below 0.002 rad, and the absolute gripper state is unchanged. Gripper transitions are always retained. Actions keep their original 10 Hz endpoint meaning and are not recomputed after filtering. Encoded source JPEGs are copied byte-for-byte into the filtered TFRecords.

• observation/image: external RGB, 224 x 224.
• observation/hand_image: wrist RGB, 224 x 224.
• observation/end_effector_pose: six UF850 joint angles in radians.
• action[0:3]: relative TCP XYZ in centimetres.
• action[3:6]: wrapped relative Euler rotation in radians.
• action[6]: absolute gripper target (-1 open, +1 closed).

Loading

from huggingface_hub import snapshot_download
import tensorflow_datasets as tfds

root = snapshot_download(repo_id="CASE-Lab/uf850-vr-teleop-hang-the-cup-rlds-noop-filtered", repo_type="dataset")
builder = tfds.builder_from_directory(f"{root}/utokyo_xarm_pick_and_place_converted_externally_to_rlds/1.0.1")
dataset = builder.as_dataset(split="train", shuffle_files=False)

Use the repository root as VLA-Adapter data_root_dir. UF850 task repositories share the same internal TFDS dataset name and must remain in separate local roots.

Provenance and limitations

SOURCE_README.md preserves the superseded source card, conversion_manifest.json records per-episode filtering and source hashes, and SHA256SUMS covers every published payload. Only a train split is provided; there are no independent per-episode success labels.