--- license: apache-2.0 task_categories: - visual-question-answering - image-to-text language: - en tags: - robotics - embodied-ai - multimodal - benchmark - vision-language - EO-1 pretty_name: EO-Bench size_categories: - n<1K --- # EO-Bench: Embodied Reasoning Benchmark for Vision-Language Models

arXiv GitHub Dataset

## Overview **EO-Bench** is a comprehensive benchmark designed to evaluate the **embodied reasoning** capabilities of vision-language models (VLMs) in robotics scenarios. This benchmark is part of the [EO-1](https://github.com/SHAILAB-IPEC/EO1) project, which develops unified embodied foundation models for general robot control. The benchmark assesses model performance across **12 distinct embodied reasoning categories**, covering trajectory reasoning, visual grounding, action reasoning, and more. All tasks are presented in a multiple-choice format to enable standardized evaluation. ## Dataset Description ### Statistics | Metric | Value | |--------|-------| | Total Samples | 600 | | Question Types | 12 | | Total Images | 668 | | Answer Format | Multiple Choice (A/B/C/D) | ### Question Type Distribution | Question Type | Count | |---------------|-------| | Trajectory Reasoning | 134 | | Visual Grounding | 128 | | Process Verification | 113 | | Multiview Pointing | 68 | | Relation Reasoning | 40 | | Robot Interaction | 37 | | Object State | 34 | | Episode Caption | 17 | | Action Reasoning | 13 | | Task Planning | 10 | | Direct Influence | 4 | | Counterfactual | 2 | ### Data Fields Each sample contains the following fields: - **`id`** (int): Unique identifier for each sample - **`question`** (str): The question text with multiple-choice options - **`question_type`** (str): Category of the question (one of 12 types) - **`answer`** (str): Correct answer letter (A, B, C, or D) - **`num_images`** (int): Number of images associated with the question - **`image_paths`** (list[str]): Relative paths to the associated images ### Question Type Descriptions | Type | Description | |------|-------------| | **Trajectory Reasoning** | Predict the optimal path for robot end-effector movement | | **Visual Grounding** | Locate specific objects or regions in the scene | | **Process Verification** | Verify the correctness of a robotic action sequence | | **Multiview Pointing** | Identify corresponding points across multiple camera views | | **Relation Reasoning** | Understand spatial relationships between objects | | **Robot Interaction** | Predict outcomes of robot-environment interactions | | **Object State** | Recognize and reason about object states | | **Episode Caption** | Describe robotic manipulation episodes | | **Action Reasoning** | Reason about the effects of robot actions | | **Task Planning** | Plan sequences of actions to achieve goals | | **Direct Influence** | Understand direct causal effects in manipulation | | **Counterfactual** | Reason about hypothetical alternative scenarios | ## Usage ### Loading the Dataset ```python from datasets import load_dataset # Load the dataset dataset = load_dataset("IPEC-COMMUNITY/EO-Bench") # Access samples for sample in dataset['train']: print(f"ID: {sample['id']}") print(f"Question: {sample['question']}") print(f"Type: {sample['question_type']}") print(f"Answer: {sample['answer']}") print(f"Images: {sample['image_paths']}") break ``` ### Loading Images ```python from PIL import Image from datasets import load_dataset import os dataset = load_dataset("IPEC-COMMUNITY/EO-Bench") # Get the first sample sample = dataset['train'][0] # Load associated images for img_path in sample['image_paths']: # Images are stored in the 'images' folder image = Image.open(hf_hub_download( repo_id="IPEC-COMMUNITY/EO-Bench", filename=img_path, repo_type="dataset" )) image.show() ``` ### Evaluation Example ```python from datasets import load_dataset def evaluate_model(model, dataset): correct = 0 total = 0 for sample in dataset['train']: # Load images and prepare input images = [load_image(p) for p in sample['image_paths']] question = sample['question'] # Get model prediction prediction = model.predict(images, question) # Check if correct if prediction == sample['answer']: correct += 1 total += 1 accuracy = correct / total * 100 return accuracy ``` ## Related Resources ### EO-1 Model EO-1 is a unified embodied foundation model that processes interleaved vision-text-action inputs using a single decoder-only transformer architecture. The model achieves state-of-the-art performance on multimodal embodied reasoning tasks. - **Paper**: [EmbodiedOneVision: Interleaved Vision-Text-Action Pretraining for General Robot Control](https://arxiv.org/abs/2508.21112) - **GitHub**: [SHAILAB-IPEC/EO1](https://github.com/SHAILAB-IPEC/EO1) - **Models**: [IPEC-COMMUNITY/EO-1-3B](https://huggingface.co/IPEC-COMMUNITY/EO-1-3B) ### Training Data EO-1 is trained on **EO-Data1.5M**, a comprehensive multimodal embodied reasoning dataset with over 1.5 million high-quality interleaved samples. - **Dataset**: [IPEC-COMMUNITY/EO-Data1.5M](https://huggingface.co/datasets/IPEC-COMMUNITY/EO-Data1.5M) ## Citation If you use this benchmark in your research, please cite: ```bibtex @article{eo1, title={EmbodiedOneVision: Interleaved Vision-Text-Action Pretraining for General Robot Control}, author={EO-1 Team}, journal={arXiv preprint arXiv:2508.21112}, year={2025} } ``` ## License This dataset is released under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0). ## Contact For questions or issues, please open an issue on the [GitHub repository](https://github.com/SHAILAB-IPEC/EO1) or contact the EO-1 team.