clean_embodied/README.md

---
license: mit
language:
- en
pretty_name: EmbodiedBench-Audio (clean speech, 300 samples)
size_categories:
- n<1K
task_categories:
- text-to-speech
- automatic-speech-recognition
- visual-question-answering
- robotics
multilinguality:
- monolingual
annotations_creators:
- machine-generated
- crowdsourced
language_creators:
- crowdsourced
source_datasets:
- extended|EmbodiedBench/EB-ALFRED
- extended|alfred
tags:
- embodiedbench
- alfred
- ai2thor
- speech
- multimodal
- omniagentbench
- embodied-agents
- household-tasks
configs:
- config_name: default
  data_files:
  - split: test
    path: clean_embodied/Embodied_task.csv
---

# Dataset Card for EmbodiedBench-Audio (clean_embodied)

## Dataset Summary

**EmbodiedBench-Audio** is the speech-modality extension of the [EmbodiedBench](https://arxiv.org/abs/2502.09560) / [EB-ALFRED](https://huggingface.co/datasets/EmbodiedBench/EB-ALFRED) benchmark. This release (`clean_embodied/`) provides 300 paired examples of `(audio, image, instruction, ground-truth plan)` covering household embodied tasks in [AI2-THOR](https://ai2thor.allenai.org/) scenes.

Each sample comes with:
- **Clean text-to-speech audio** of the natural-language goal
- The **starting-frame screenshot** the agent sees
- The **original written goal** and **ground-truth high-level plan** (both natural-language step descriptions and symbolic PDDL actions)

The folder is named `clean_embodied/` so contributors can drop sibling folders like `noisy_embodied_<env>/` later without overwriting the clean baseline.

## Supported Tasks

- **Embodied speech grounding** — predict the action plan from spoken goal + image
- **Speech-to-PDDL plan generation** — speech input → symbolic action sequence
- **ASR robustness** — comparing models that transcribe-then-plan vs. models that ingest audio directly
- **Cross-modal evaluation** — audio vs. text input on identical samples

## Languages

English (en, US accent, single TTS voice).

## Dataset Structure

### Layout

```
clean_embodied/
├── README.md                          (this card)
├── Embodied_task.csv                  300 rows · 7 columns — source of truth
├── audio/
│   ├── manifest.json                  sample_id → all CSV fields + audio_file + spoken_text
│   └── embodied_<trial>_<repeat_idx>.wav   300 files · 24 kHz mono · ~35 MB
└── images/
    └── <trial>_<repeat_idx>.png       300 files · ~33 MB · AI2-THOR scene capture
```

### Data Fields (Embodied_task.csv)

| Field | Type | Description |
|---|---|---|
| `trial` | string | Original ALFRED trial id, e.g. `trial_T20190908_044113_026049` |
| `repeat_idx` | int | Annotator index; each trial has 3+ paraphrasings of the same goal by different turkers |
| `scene` | string | AI2-THOR floor plan, e.g. `FloorPlan219` |
| `goal_instr` | string | One-sentence high-level goal — the **input text** the agent receives |
| `gt_high_descs` | list[string] (JSON-encoded) | Step-by-step natural-language descriptions — auxiliary annotation |
| `gt_high_pddl_actions` | list[string] (JSON-encoded) | Symbolic ground-truth plan, e.g. `["GotoLocation ['dresser']", "PickupObject ['cellphone']", ...]` |
| `image_name` | string | Filename of the starting frame in `images/` |

### Data Fields (audio/manifest.json)

For every sample id, the manifest records:

| Key | Description |
|---|---|
| `sample_id` | `embodied_<trial>_<repeat_idx>` |
| `audio_file` | `<sample_id>.wav` |
| `spoken_text` | Identical to `goal_instr` — the text fed to the TTS model |
| `trial`, `repeat_idx`, `scene`, `image_name`, `gt_high_descs`, `gt_high_pddl_actions` | Mirrored from CSV for convenience |

### Pairing across modalities

```python
sample_id  = f"embodied_{trial}_{repeat_idx}"
audio_path = f"clean_embodied/audio/{sample_id}.wav"
image_path = f"clean_embodied/images/{trial}_{repeat_idx}.png"
```

### Splits

A single `test` split with 300 examples. The CSV is a curated subset of EB-ALFRED's `valid_seen` / `valid_unseen` splits sampled across multiple ALFRED task types (e.g. `look_at_obj_in_light`, `pick_and_place_simple`, `pick_clean_then_place_in_recep`, `pick_heat_then_place_in_recep`, `pick_cool_then_place_in_recep`).

## Dataset Creation

### Curation Rationale

EmbodiedBench evaluates language-conditioned embodied agents on text + image, but the original benchmark has no audio modality. This release adds one: each goal instruction is rendered as clean speech so models with native audio input (Qwen2.5-Omni, Qwen3-Omni, etc.) can be evaluated in the same protocol. The 300-sample subset balances coverage across ALFRED task types while staying small enough for fast iteration.

### Source Data

- **Goal instructions, step descriptions, PDDL plans:** [EB-ALFRED](https://huggingface.co/datasets/EmbodiedBench/EB-ALFRED), which itself is a subset of [ALFRED](https://askforalfred.com/) (Shridhar et al., CVPR 2020). The natural-language fields were collected via Amazon Mechanical Turk on ALFRED trajectories.
- **Images:** Starting frames extracted from AI2-THOR scene trajectories included in EB-ALFRED.
- **Subset selection:** 300 trial/annotator pairs sampled by the OmniAgentBench team to span ALFRED task types.

### Annotations

- **Audio:** Machine-generated. TTS via [`Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign`](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign) on a single NVIDIA RTX PRO 6000 Blackwell GPU. Voice instruct prompt: *"Speak naturally and clearly as if giving a household task instruction to a robot assistant."* Output: 24 kHz mono WAV, bf16 inference, no flash-attn. Generated 2026-05-03; ~10 min wall time for 300 samples.
- **All other fields** (text, plans, scene, image): inherited unmodified from EB-ALFRED / ALFRED.

### Personal and Sensitive Information

None. All instructions describe household task goals with no identifiable people, addresses, or PII.

## Audio Characteristics

- **Sample rate / channels:** 24 kHz · mono
- **Duration:** 1.2 s – 5.1 s per WAV (median 2.3 s, mean 2.5 s)
- **Speech rate:** ~208 wpm — fast, because source text is short (median 9 words)
- **Voice:** Single Qwen3-TTS-VoiceDesign voice across all 300 samples — no speaker variation in this clean release

## Considerations for Using the Data

### Limitations

- **One TTS voice, no speaker variation.** All 300 samples use the same synthetic voice. Results should not be interpreted as covering speaker diversity.
- **Short audio.** Goal instructions are one-sentence (median 9 words), so audio is brief. Models that benefit from longer context windows may not fully exercise their capacity here.
- **Speech rate is fast.** ~208 wpm vs. natural English ~150 wpm. The TTS does not include expressive prosody beyond what the voice-design instruct prompt nudges.
- **Audio is fully synthetic.** No human-recorded speech. Add `noisy_*/` siblings (background noise, reverb, accents, real recordings) to extend robustness coverage.

### Biases & Social Impact

The underlying ALFRED instructions describe everyday household chores in stereotypical kitchen/bedroom/bathroom layouts and reflect the cultural framing of the original Mechanical Turk annotators. Models trained or evaluated on this data will inherit those framings.

### Other Known Issues

- The 300-sample subset is a **convenience sample**, not a stratified random sample, and is not balanced for AI2-THOR scene type or PDDL plan length.

## Additional Information

### Dataset Curators

OmniAgentBench team (audio synthesis, packaging, curation of the 300-sample subset). Upstream natural-language and trajectory data: EmbodiedBench team (Yang et al., 2025) and ALFRED team (Shridhar et al., 2020).

### Licensing Information

Released under MIT, **subject to the licenses of the upstream datasets**:
- [EmbodiedBench](https://github.com/EmbodiedBench/EmbodiedBench) — see upstream repo
- [ALFRED](https://github.com/askforalfred/alfred) — MIT
- [AI2-THOR](https://github.com/allenai/ai2thor) (used to render the images) — Apache 2.0

If you redistribute, retain attribution to all three upstream sources.

### Citation Information

```bibtex
@misc{omniagentbench_embodied_audio_2026,
  title  = {EmbodiedBench-Audio: clean speech for the EmbodiedBench 300-sample subset},
  author = {OmniAgentBench Team},
  year   = {2026},
  url    = {https://huggingface.co/datasets/omniagentbench/EmbodiedBench-Audio}
}

@inproceedings{yang2025embodiedbench,
  title  = {EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents},
  author = {Yang, Rui and others},
  year   = {2025},
  eprint = {2502.09560},
  archivePrefix = {arXiv}
}

@inproceedings{shridhar2020alfred,
  title     = {ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks},
  author    = {Shridhar, Mohit and Thomason, Jesse and Gordon, Daniel and Bisk, Yonatan and Han, Winson and Mottaghi, Roozbeh and Zettlemoyer, Luke and Fox, Dieter},
  booktitle = {CVPR},
  year      = {2020}
}
```

### Contributions

Contributions of `noisy_<env>/`, `wild_<variant>/`, or larger-scale follow-ups are welcome — the repo is intentionally laid out so new sibling folders sit cleanly alongside `clean_embodied/`. Open a discussion or PR on the [HuggingFace dataset page](https://huggingface.co/datasets/omniagentbench/EmbodiedBench-Audio).

Part of the [OmniAgentBench](https://huggingface.co/datasets/omniagentbench/OmniAgentBench) project.