AAyano/uf850-vr-teleop-press-buttons-rlds-noop-filtered

UF850 Quest VR Press Buttons — RLDS No-Op Filtered

OpenVLA-OFT-compatible RLDS/TFDS trajectories derived from the validated UF850 Press Buttons raw dataset. Idle steps were removed before image processing; retained actions keep their original adjacent-sample meaning and are not recomputed after filtering.

Task instruction

Every step uses the same English instruction:

First press the blue button, then press the red one, and finally press the orange one

Summary

Item	Value
Episodes	50
Raw transition steps	20,659
Retained RLDS steps	10,589
Removed no-op steps	10,070
Retained fraction	51.26%
Split	train only
TFDS version	1.0.1

The source revision is 495ebc9ff6bf301cccdcfe6765539bce4b9533e7. The raw dataset card explains the earlier rejection of one invalid capture; this released task therefore contains episode_000 through episode_049.

No-op filter

A step is removed only when all three conditions hold:

• translation norm is below 0.02 cm/step;
• wrapped rotation norm is below 0.002 rad/step;
• the binary gripper state is unchanged from the preceding raw step.

Gripper transitions are always retained. No episode required the fewer-than-two steps fallback. Filtering is performed from JSON first, so images belonging to removed steps are never cropped or encoded.

OFT transforms and schema

• steps/observation/image: external cam_1.jpg, cropped x=540:1620, y=0:1080, then LANCZOS-resized to 224×224 RGB.
• steps/observation/hand_image: wrist cam_0.jpg, cropped x=760:1840, y=0:1080, then LANCZOS-resized to 224×224 RGB.
• steps/observation/end_effector_pose: six joint angles in radians. The key name is retained solely for compatibility with the existing XArm OFT config; these values are not Cartesian end-effector poses.
• steps/action: seven float32 values: TCP XYZ delta in centimetres, wrapped Euler delta in radians, and binary gripper state.
• gripper action: +1 closed and -1 open. Each episode starts open (-1); action.delta_gripper < -10 switches to closed, > 10 switches to open, and values in between retain the previous state. This matches the legacy setting1/setting2 conversion.
• discount=1; only the final retained step has reward=1, is_last=true, and is_terminal=true.
• language_instruction contains the prompt above on every retained step.

The button task intentionally uses a closed gripper for nearly all retained motion. This is a property of the demonstrations, not a conversion error.

Layout and loading

README.md
conversion_manifest.json
SHA256SUMS
utokyo_xarm_pick_and_place_converted_externally_to_rlds/
  1.0.1/
    dataset_info.json
    features.json
    *-train.tfrecord-*

Download and open the generated builder directly:

hf download AAyano/uf850-vr-teleop-press-buttons-rlds-noop-filtered \
  --repo-type dataset --local-dir ./press_buttons_rlds

import tensorflow_datasets as tfds

builder = tfds.builder_from_directory(
    "./press_buttons_rlds/"
    "utokyo_xarm_pick_and_place_converted_externally_to_rlds/1.0.1"
)
dataset = builder.as_dataset(split="train", shuffle_files=False)

For the local OpenVLA-OFT loader, use the downloaded repository root as data_root_dir and utokyo_xarm_pick_and_place_converted_externally_to_rlds as dataset_name. This dataset was generated and validated with TensorFlow 2.15.0 and TFDS 4.9.3.

The corn and button releases use the same internal TFDS name, so download them into separate parent directories. Do not merge their version directories.

Reproducibility and validation

conversion_manifest.json records every episode's source JSON digest, raw and retained counts, action statistics, preprocessing parameters, runtime versions, and converter digest. SHA256SUMS covers the published payload. The completed RLDS was reopened in a fresh read-only builder and every retained state, action, prompt, flag, and decoded image was checked.