RobotEQ3.0

RobotEQ3.0 is a multimodal benchmark for personalized, multi-label robot behavior preference prediction. Given a scenario image, a robot-role question, a set of candidate actions, and an anonymized participant profile, the task is to predict all actions preferred by that participant.

Dataset Viewer

The default Viewer is backed by image-embedded Parquet shards and contains one row per scenario, displaying the scene image, scenario ID, question, and candidate action list. Ground-truth participant preferences are provided separately in annotations/preference_annotations.json.

Dataset Statistics

Item	Count
Scenarios	1,800
Images	1,800
Anonymous participants	10
Profile dimensions	9
Participant-scenario annotations	18,000
Executable candidate actions per scenario	5
No-clear-preference label	F

Task

For participant $u$ and scenario $i$, a model receives:

• a scenario image;
• a question describing the robot role;
• five executable candidate actions (A–E);
• an anonymized nine-dimensional participant profile.

The model predicts a set of preferred action labels. Since participants may select multiple actions, this is a multi-label prediction task. Label F indicates that the participant had no clear preference because the listed options were similarly appropriate.

Repository Structure

RobotEQ3.0/
├── README.md
├── metadata.json
├── viewer.jsonl
├── viewer/                 # Hugging Face Viewer Parquet shards
├── scenarios/
│   ├── scenarios.json
│   └── images/
│       ├── 00001.jpg
│       ├── 00002.jpg
│       └── ...
├── participants/
│   ├── participant_profiles.json
│   └── profile_schema.json
└── annotations/
    └── preference_annotations.json

Scenario Format

{
  "scenario_id": "00001",
  "image_path": "images/00001.jpg",
  "question": "You are a subway station attendant robot. What is the most appropriate action to take next?",
  "actions": {
    "A": "...",
    "B": "...",
    "C": "...",
    "D": "...",
    "E": "...",
    "F": "No clear preference; the listed options are similarly appropriate"
  }
}

Participant Profiles

Participants are represented only by anonymous identifiers P1–P10. Each profile contains nine aggregate dimensions:

Code	Dimension	Range
RT	Robot Trust	1–7
O	Openness	1–5
C	Conscientiousness	1–5
E	Extraversion	1–5
A	Agreeableness	1–5
N	Neuroticism	1–5
AU	Autonomy	1–5
PT	Perspective Taking	0–4
EC	Empathic Concern	0–4

Preference Annotation Format

{
  "participant_id": "P1",
  "scenario_id": "00001",
  "selected_actions": ["A", "C"]
}

selected_actions is always represented as a list. Three original records with no recorded selection are preserved as empty lists rather than imputed.

Loading

import json
from pathlib import Path

root = Path("RobotEQ3.0")
scenarios = json.loads((root / "scenarios/scenarios.json").read_text(encoding="utf-8"))
profiles = json.loads((root / "participants/participant_profiles.json").read_text(encoding="utf-8"))
annotations = json.loads((root / "annotations/preference_annotations.json").read_text(encoding="utf-8"))

The Viewer subset can also be loaded directly from the Hub:

from datasets import load_dataset

dataset = load_dataset("Tongji-Emotion/RobotEQ3.0", split="train")

Privacy

The public release excludes real names, contact information, identity mappings, raw item-level questionnaire responses, free-text participant feedback, and model predictions. Participant IDs are anonymous and cannot be mapped to real identities from the released files.

License

A redistribution license has not yet been declared in the Dataset Card. Users should consult the repository maintainers before redistribution or commercial use.