Datasets:

EMBGuard
/

EMBGuardTest_v3

---
license: apache-2.0
language:
- en
pretty_name: EMBGuardTest v3
size_categories:
- n<1K
task_categories:
- image-text-to-text
tags:
- embodied-ai
- safety
- guardrail
- physical-hazard
- benchmark
configs:
- config_name: default
  data_files:
  - split: causal_risky
    path: data/causal_risky-*
  - split: decoupled_benign
    path: data/decoupled_benign-*
  - split: selective_risky
    path: data/selective_risky-*
  - split: absent_benign
    path: data/absent_benign-*
dataset_info:
  features:
  - name: id
    dtype: string
  - name: image
    dtype: image
  - name: image_path
    dtype: string
  - name: risk
    dtype: string
  - name: type
    dtype: string
  - name: scenario_type
    dtype: string
  - name: risk_type
    dtype: string
  - name: risk_subtype
    dtype: string
  - name: situation
    dtype: string
  - name: hazard
    dtype: string
  - name: action
    dtype: string
  - name: room
    dtype: string
  - name: multi_scenario
    dtype: string
  - name: pair_item_id
    dtype: string
  splits:
  - name: causal_risky
    num_bytes: 264499562
    num_examples: 75
  - name: decoupled_benign
    num_bytes: 264500374
    num_examples: 75
  - name: selective_risky
    num_bytes: 401910525
    num_examples: 104
  - name: absent_benign
    num_bytes: 263962112
    num_examples: 75
  download_size: 1194900005
  dataset_size: 1194872573
---

# EMBGuardTest v3

EMBGuardTest is the evaluation set of **EMBGuard** ("EMBGuard: Constructing Hazard-Aware Guardrails for Safe Planning in Embodied Agents", [arXiv:2605.30924](https://arxiv.org/abs/2605.30924)).
Each example is a real-world indoor photograph plus a proposed robot action. A guardrail model must decide whether
executing the action in that scene is **safe** or **unsafe**.

v3 supersedes `EMBGuard/EMBGuardTest_v2`: some images and ids in v2 were wrong (mismatched or duplicated images, swapped
selective pairs, inconsistent ids), and the id scheme changed. **Results on v2 and v3 are not directly comparable.**

## Splits

| split | rows | label (`risk` / `type`) | meaning |
|---|---|---|---|
| `causal_risky` | 75 | `O` / `unsafe` | A hazard is present and causally relevant to the action. |
| `decoupled_benign` | 75 | `X` / `safe` | Same image as `causal_risky`; a hazard-like element is present but decoupled from the action, so the scene should be judged safe. |
| `selective_risky` | 104 | `O` / `unsafe` | Several hazard candidates are present; the model must identify the action-relevant risky condition. 52 images, each used by exactly 2 rows. |
| `absent_benign` | 75 | `X` / `safe` | The hazard is absent from the scene (same scenario number as `causal_risky`). |

The 75 scenarios in `causal_risky`, `decoupled_benign` and `absent_benign` share numbers: `causal_risky-022`,
`decoupled_benign-022` and `absent_benign-022` are the same scenario. `decoupled_benign` re-uses the `causal_risky` image.

## Columns

| column | type | description |
|---|---|---|
| `id` | string | `{split}-{NNN}`, e.g. `causal_risky-022`. Unique within a split; not meaningful for inference. |
| `image` | image | JPEG photograph (EXIF orientation applied; resolution varies, long side up to 5712 px). |
| `image_path` | string | Relative path used while building the dataset (informational; files are embedded in the parquet). |
| `risk` | string | `O` (unsafe) or `X` (safe). |
| `type` | string | `unsafe` or `safe`. |
| `scenario_type` | string | One of the four split names. |
| `risk_type` | string | One of 7 hazard categories: `Electrical_Risk`, `Fire_Risk`, `Crush_Pinch_Risk`, `Cut_Sharp_Risk`, `Contamination_Infection_Risk`, `Slip_Trip_Fall_Risk`, `Chemical_Toxic_Exposure_Risk`. |
| `risk_subtype` | string | Finer hazard category (23 values in the causal scenarios, 19 in `selective_risky`). |
| `situation` | string | Unused, always empty (kept for schema compatibility with v2). |
| `hazard` | string | Short description of the hazard in the scene. For `absent_benign` it describes the hazard-free scene. |
| `action` | string | The proposed robot action to be judged. |
| `room` | string | Room type of the scene: `Kitchen`, `Livingroom`, `Bedroom`, `Bathroom`, `Office` or `Hallway`. Filled for every row. The dataset authors labelled the 48 non-augmented `selective_risky` rows; the other rows were labelled afterwards by manual review of the photographs. |
| `multi_scenario` | string | `"yes"` for the 56 `selective_risky` rows whose image contains two scenarios, `"no"` for all other rows. |
| `pair_item_id` | string | `selective_risky` only: `id` of the partner row that shares the same image (`selective_risky-000` <-> `selective_risky-001`). Empty in other splits. |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("EMBGuard/EMBGuardTest_v3")
ex = ds["causal_risky"][0]
print(ex["id"], ex["action"], ex["hazard"])
ex["image"]  # PIL.Image
```

Evaluation scripts and the guardrail models are in the [EMBGuard GitHub repository](https://github.com/dongwxxkchoi/EMBGuard).

## Notes and limitations

- Small by design (329 rows): intended for evaluation only, not for training.
- 27 images have a short side below 1000 px (down to 360 px).
- `selective_risky` rows are only meaningful together with their partner row (same image, different hazard/action).

## Citation

```bibtex
@article{choi2026embguard,
  title   = {EMBGuard: Constructing Hazard-Aware Guardrails for Safe Planning in Embodied Agents},
  author  = {Choi, Dongwook and Kwon, Taeyoon and Jeong, Bogyung and Kim, Minju and Hwang, Yeonjun and Kim, Hyojun and Kim, Byungchul and Jang, Young Kyun and Yeo, Jinyoung},
  journal = {arXiv preprint arXiv:2605.30924},
  year    = {2026},
  url     = {https://arxiv.org/abs/2605.30924}
}
```