README.md

---
license: cc-by-4.0
size_categories:
- 1K<n<10K
pretty_name: BEHAVIOR ESI-Bench
task_categories:
- robotics
tags:
- json
- text
- croissant
- pandas
- polars
- region:us
configs:
- config_name: default
  data_files:
  - split: train
    path: data/questions.jsonl
---

# BEHAVIOR ESI-Bench

[**Project Page**](https://esi-bench.github.io/) | [**Paper**](https://huggingface.co/papers/2605.18746) | [**GitHub**](https://github.com/ESI-Bench/ESI-Bench)

BEHAVIOR ESI-Bench is a comprehensive benchmark for evaluating embodied spatial intelligence across indoor scenes, object arrangements, physical reasoning, temporal understanding, and active exploration tasks. Built on OmniGibson and grounded in Spelke's core knowledge systems, it spans 10 task categories and 29 subcategories. 

Unlike passive sensing benchmarks, ESI-Bench requires agents to close the perception-action loop: deciding what abilities to deploy—perception, locomotion, and manipulation—and how to sequence them to actively accumulate task-relevant evidence.

## Schema

Each row is one question instance. The table is intentionally flattened for Hugging Face Dataset Viewer and Croissant compatibility.

```text
id
big_task
small_task
runner_task
scene
room
question
answer
answer_type
options_json
image_paths_json
metadata_json
```

`options_json`, `image_paths_json`, and `metadata_json` are JSON-encoded strings. They can be decoded by downstream users when structured fields are needed.

## Task Taxonomy

The dataset follows the ESI-Bench table hierarchy:

```text
Action Sequencing
Cognitive Mapping
Enumerative Perception
Metric Comparison
Perceptual Grounding
Physical Dynamics
Physical Structure
Spatial Relations
Specular Reflection
Temporal Understanding
```

The `small_task` column stores the corresponding subtask, and `runner_task` stores the internal task module name used by the original BEHAVIOR active-exploration code.

## Sample Usage

To run the active exploration module using a task from the dataset, you can use the provided runner script in the [official repository](https://github.com/ESI-Bench/ESI-Bench):

```bash
# Example: Running the Explorer for a counting task using the Gemini provider
python src/main.py \
  --task counting \
  --metadata "dataset/json_clean/Enumerative Perception/Spatial Segmentation/Merom_0_int/living_room_0/q_000.json" \
  --provider gemini \
  --model gemini-3.1-pro-preview \
  --max-steps 30 \
  --min-steps 1 \
  --threshold 0.9 \
  --results-root outputs/results \
  --step-image-root outputs/steps \
  --overwrite
```

## Citation

If you find ESI-Bench useful in your research, please cite:

```bibtex
@inproceedings{hong2026esibench,
  title     = {{ESI-Bench}: Towards Embodied Spatial Intelligence that Closes the Perception-Action Loop},
  author    = {Hong, Yining and Liu, Jiageng and Yin, Han and Li, Manling and Guibas, Leonidas and Li, Fei-Fei and Wu, Jiajun and Choi, Yejin},
  year      = {2026}
}
```

This benchmark is built on [BEHAVIOR-1K](https://behavior.stanford.edu/behavior-1k/) and [OmniGibson](https://behavior.stanford.edu/omnigibson/). Please consider citing those works as well if you use this dataset.

## Croissant

Hugging Face automatically generates Croissant metadata from the Dataset Viewer once this dataset is processed. The Croissant JSON-LD endpoint is:

```text
https://huggingface.co/api/datasets/ESI-Bench/esi-bench/croissant
```