iciclab/Robot_physics_interaction_with_Human_linguistic

Dual-Camera Speech and Interaction Dataset

This directory contains the processed output of 50 synchronized recording sessions. Each session combines first-person video, third-person video, human speech, 30 fps image sequences, and a text label derived from the read-aloud script.

For the research context behind the dataset, see PROJECT_README.md.

Directory structure

ACTION_NUMBER/
├── first_person_view.mp4
├── third_person_view.mp4
├── audio.wav
├── text.txt
├── label.json
├── first_person_view_camera/
│   ├── first_person_view_camera_1.jpg
│   ├── first_person_view_camera_2.jpg
│   └── ...
└── third_person_view_camera/
    ├── third_person_view_camera_1.jpg
    ├── third_person_view_camera_2.jpg
    └── ...

Files in each session

File or directory	Description
first_person_view.mp4	First-person video, 640 x 480, approximately 10 fps; the source image was rotated 180 degrees during recording
third_person_view.mp4	Third-person video, 424 x 240, approximately 5 fps
audio.wav	Corresponding two-channel human speech recording at 48 kHz
first_person_view_camera/*.jpg	First-person video converted to a 30 fps JPEG sequence
third_person_view_camera/*.jpg	Third-person video converted to a 30 fps JPEG sequence
text.txt	Clean script text associated with the session
label.json	Script identity, marked and clean text, ASR output, match scores, and review status

Frame numbering starts at 1. For either camera, frame N corresponds to (N - 1) / 30 seconds from the beginning of that camera stream.

Session naming

Processed session directories are numbered sequentially within each physical action:

press_01
press_02
pull_01
slide_01
twist_01
insert_01

The first component is the physical action, and the integer is the recording's chronological position within that action. Numbering starts at 1 for each action. Prosody and script information are retained in scenario_id and prosody_profile inside the label files rather than encoded in the directory name. The original timestamp directory name remains available as source_session.

Frame-rate conversion

The original first-person and third-person videos were recorded at approximately 10 fps and 5 fps, respectively. The image sequences were generated with the FFmpeg fps=30 filter. This operation repeats the nearest source frame on a 30 Hz timeline; it does not perform optical-flow or generative interpolation. Adjacent JPEG files may therefore contain the same source image.

The two cameras can have slightly different start and end delays. As a result, the first-person and third-person image counts within one session may differ. Each stream has its own frame index and timestamp beginning at zero.

Frame manifest

frame_manifest.csv contains one row per image and can be used directly to construct image-text samples.

Field	Description
session	Action-based processed directory name, such as press_01
camera	first_person_view_camera or third_person_view_camera
frame_index	One-based frame number
timestamp_s	Time from the beginning of the camera stream in seconds
image_path	Image path relative to this directory
scenario_id	Script identifier, such as PRES-U01
text	Clean script text associated with the session

Every image from the same session shares that session's utterance-level text label.

Session labels

label.json, session_labels.json, and session_labels.csv use the following main fields:

Field	Description
source_session	Original timestamp directory name, such as session_20260918_120529
session	Action-based processed directory name
scenario_id	Script identifier
title	Script-item title
task	Interaction type: PRES, PULL, SLID, TWST, or INSR
prosody_profile	Prosody class: U, E, C, or R
marked_text	Original script with performance directions, stress, and pause notation
text	Clean text used as the image label
asr_transcript	Whisper transcription of the recorded speech
match_score	Similarity between the ASR transcript and the selected script, from 0 to 1
match_margin	Difference between the best and second-best automatic match scores
match_method	asr_fuzzy_match or sequence_override
needs_review	true when the recording should be checked manually