🧠 J-ORA: A Robot Perception Framework for Japanese Object Identification, Reference Resolution, and Action Prediction

A Multimodal Dataset for Vision-Language Grounding in Human-Robot Interaction (HRI)
📍 Language: Japanese | 🤖 Focus: Embodied AI | 📆 Size: 142 Scenes | 🔬 Granularity: Object-Level

---

📘 Summary

J-ORA (Japanese Object Reference and Action) is a multimodal benchmark for grounded vision-language learning in robotics. It is designed for understanding Japanese robot instructions in real-world settings through a combination of:

• Rich object-level visual scenes
• Human-robot dialogues in Japanese
• Grounded object attributes and reference links
• Support for 3 robot perception tasks:
  • Object Identification (OI)
  • Reference Resolution (RR)
  • Next Action Prediction (AP)

---

🌏 Motivation

Despite the growing role of Vision-Language Models (VLMs), robot perception remains challenging in dynamic real-world environments with occlusions, multiple object types, and ambiguous language.
Yet, existing benchmarks focus mainly on English and synthetic prompts.
J-ORA fills this gap with a Japanese-centric, multimodal resource grounded in real images and conversational instructions, supporting the development of intelligent agents that understand and act in non-English human environments.

---

📦 Dataset Overview

Feature	Value
Hours of recordings	3 hrs 3 min 44 sec
Unique dialogues	93
Total dialogue scenes	142
Total utterances	2,131
Average turns per dialogue	15
Image-dialogue pairs	142
Unique object classes	160
Object attribute annotations	1,817
Language	Japanese

Each .png image is paired with a .json file that contains:

• 🔨 Full Japanese dialogue text
• 🔍 Object attributes (category, color, shape, size, material, etc.)
• 📍 Reference links from language to visual regions
• ⚙️ Spatial positions and interactivity annotations

---

📂 Structure

Each sample contains:

{
  "image_id": "001",
  "caption": {
    "text": "Japanese instruction dialogue..."
  },
  "i2t_relations": [...],
  "object_attributes": [...],
  "relations": [...]
}

Files:

001.png
001.png.json
002.png
002.png.json
...

To use:

from datasets import load_dataset

ds = load_dataset("jatuhurrra/J-ORA", split="train")

---

🧠 Tasks Supported

🟡 Object Identification (OI)

Identify all object mentions in a Japanese instruction and match them to objects in the scene.

🔵 Reference Resolution (RR)

Locate the visual regions of mentioned objects based on textual referring expressions.

🔴 Action Prediction (AP)

Given the mentioned object(s) and their locations, predict the likely action (e.g., pick, move, discard).

---

🧪 Evaluations

We benchmarked 10+ VLMs, including:

• 🏆 Proprietary: GPT-4o, Gemini 1.5 Pro, Claude 3.5
• 💻 General open-source: LLaVa, Qwen2VL
• 🇯🇵 Japanese-specific: EvoVLM-JP, Bilingual-gpt-neox, Japanese-stable-VLM

We test zero-shot and fine-tuned variants with and without object attribute embeddings.
Findings reveal:

• Performance gaps persist between English and Japanese
• Reference resolution remains the hardest task
• Attribute-based fine-tuning improves grounding

---

🔍 Use Cases

J-ORA supports research in:

• Japanese robot instruction following
• Multilingual grounding and reference resolution
• Vision-language model benchmarking
• Embodied AI with language grounding
• Fine-tuning of Japanese VLMs for HRI tasks

---

🛠️ Resources

• Code: Task pipelines, training scripts, and evaluation metrics.

• Dataset: Full annotations and image data.

  The data introduced in this project extends the J-CRe3 data.

---

📄 Citation

If you use this dataset, please cite both J-ORA and the J-CRe3 dataset:

% J-ORA citation coming soon

@inproceedings{ueda-2024-j-cre3,
  title     = {J-CRe3: A Japanese Conversation Dataset for Real-world Reference Resolution},
  author    = {Nobuhiro Ueda and Hideko Habe and Yoko Matsui and Akishige Yuguchi and Seiya Kawano and Yasutomo Kawanishi and Sadao Kurohashi and Koichiro Yoshino},
  booktitle = {Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
  month     = may,
  year      = {2024},
  url       = {https://aclanthology.org/2024.lrec-main.829},
  pages     = {9489--9502},
  address   = {Turin, Italy},
}

---

📬 Contact

For questions and collaborations:

• Jesse Atuhurra — atuhurra.jesse.ag2@naist.ac.jp
• Koichiro Yoshino — koichiro.yoshino@riken.jp

---

📜 License

Released under:
CC BY-SA 4.0 — Creative Commons Attribution-ShareAlike 4.0 International License.

---