LeoJiangOR/opentouch-vitra-stage1-keypoint-full

OpenTouch VITRA Stage-1 Keypoint Conversion

This repository contains a VITRA-compatible conversion of the locally downloaded OpenTouch dataset. It is intended for experiments on tactile-aware hand action prediction, tactile editing, and test-time guidance with VITRA-style human action models.

This is a derived dataset conversion. Users should follow the terms of the original OpenTouch dataset when using these files.

Dataset Size

The converted dataset was built from all OpenTouch raw HDF5 files available locally at conversion time.

Item	Count
Raw HDF5 files processed	26
OpenTouch clips seen	2,958
VITRA episodes written	2,958
Total frames written	327,030
Skipped short clips	0
Skipped low-valid-ratio clips	0
Conversion errors	0

Split summary:

Split	Episodes	Frame entries
opentouch_keypoint_train	2,366	261,211
opentouch_keypoint_test	592	65,819

Directory Layout

Annotation/
  opentouch_keypoint_train/
    episode_frame_index.npz
    episodic_annotations/*.npy
    conversion_report.json
  opentouch_keypoint_test/
    episode_frame_index.npz
    episodic_annotations/*.npy
    conversion_report.json
  statistics/
    opentouch_keypoint_train_keypoints_statistics.json
Video/
  OpenTouch_root/
    <session>/<demo>.mp4
conversion_report.json

Each episodic_annotations/*.npy file stores a Python dictionary in the VITRA Stage-1 episodic format.

What Was Converted

OpenTouch raw HDF5 clips do not provide a ready-made VLA action vector. The raw clips contain synchronized streams such as:

rgb_images_jpeg:        [T]
timestamps:             [T]
camera_poses:           [T, 4, 4]
right_hand_landmarks:   [T, 21, 3]
right_palm_pos:         [T, 3]
right_wrist_pos:        [T, 3]
right_pressure:         [T, 16, 16]

The accompanying CSV annotations contain semantic labels and language fields such as object name, object category, action label, grip type, and a natural-language description. Those fields are used to build the VITRA text instruction.

Action Target Construction

The low-level VITRA action target is derived from the future trajectory of the OpenTouch 3D hand landmarks.

For each frame, the converter stores right-hand landmarks as VITRA-style hand state fields:

joints_worldspace: [T, 21, 3]
joints_camspace:   [T, 21, 3]
transl_worldspace: [T, 3]
transl_camspace:   [T, 3]
kept_frames:       [T]

For VITRA training with action_type="keypoints", the dataloader builds each per-step hand action as:

3 dims: relative wrist translation from t to t+1
3 dims: relative wrist rotation from t to t+1
63 dims: next-frame 21 x 3 root-relative hand keypoints

Thus one hand has 69 action dimensions. Two hands have 138 dimensions before padding.

The converted OpenTouch data is right-hand supervised. If left-hand landmarks are missing, the left hand is filled with dummy zeros and masked as invalid.

Why Pad to VITRA Dimensions

VITRA's human action interface uses fixed tensor dimensions:

state_dim = 212
action_dim = 192
horizon = 16

OpenTouch keypoint actions are smaller than this unified interface. The dataloader pads actions and states to VITRA's expected dimensions so the same VITRA model, normalizer, diffusion action head, and evaluation code can be reused across datasets.

Padding does not create new supervision. The padded dimensions are zero-filled and excluded by the action/state masks.

Tactile Fields

OpenTouch pressure maps are preserved under the episode's opentouch payload:

touch_pressure:           [T, 2, 16, 16]
touch_mask:               [T, 2]
touch_timestamps:         [T_touch]
touch_aligned_indices:    [T]
touch_aligned_timestamps: [T]
touch_alignment_valid:    [T]

The channel order is:

0: left tactile sensor
1: right tactile sensor

For most OpenTouch clips used here, the right channel is supervised and the left channel may be empty.

How to Use with VITRA-HAND

Place or download this dataset under a path such as:

datasets/vitra_opentouch_keypoint_full

Then point the VITRA-HAND dataloader to:

video_root:      datasets/vitra_opentouch_keypoint_full/Video
annotation_file: datasets/vitra_opentouch_keypoint_full/Annotation/opentouch_keypoint_train/episode_frame_index.npz
label_folder:    datasets/vitra_opentouch_keypoint_full/Annotation/opentouch_keypoint_train/episodic_annotations
statistics_path: datasets/vitra_opentouch_keypoint_full/Annotation/statistics/opentouch_keypoint_train_keypoints_statistics.json
action_type:     keypoints

For evaluation, use the opentouch_keypoint_test annotation split.

Important Notes

• The raw OpenTouch semantic action field is a language/task label, not the continuous VITRA action vector.
• The continuous action target used by VITRA is derived from future 3D hand landmark trajectories.
• This conversion is designed for keypoint-action prediction, not MANO-pose prediction.
• The converted RGB videos were decoded from OpenTouch rgb_images_jpeg frames.
• The test split does not include a separate statistics file; use the train statistics for normalization, consistent with standard train/test evaluation practice.