oscarqjh/EB-Alfred_easi

EB-ALFRED Dataset

EB-ALFRED is a benchmark dataset for evaluating embodied AI agents on household tasks. It is part of the EmbodiedBench benchmark suite.

Dataset Description

This dataset contains task specifications and instructions for the ALFRED (Action Learning From Realistic Environments and Directives) simulation environment. The tasks involve common household activities like picking up objects, placing them in different locations, cleaning, heating, and cooling items.

Subsets

The dataset is organized into 6 subsets, each testing different capabilities:

Subset	Description	Examples
base	Standard household tasks	50
common_sense	Tasks requiring common sense reasoning	50
complex_instruction	Tasks with complex natural language instructions	50
long_horizon	Multi-step tasks requiring long-term planning	50
spatial	Tasks requiring spatial reasoning	50
visual_appearance	Tasks requiring visual understanding	50

Task Types

The dataset includes 7 task types:

• pick_and_place_simple - Pick up an object and place it somewhere
• pick_and_place_with_movable_recep - Pick up an object with a container
• pick_clean_then_place_in_recep - Clean an object and place it
• pick_cool_then_place_in_recep - Cool an object and place it
• pick_heat_then_place_in_recep - Heat an object and place it
• pick_two_obj_and_place - Pick up two objects and place them
• look_at_obj_in_light - Examine an object under a light source

Dataset Structure

.
├── data/
│   ├── base.jsonl
│   ├── common_sense.jsonl
│   ├── complex_instruction.jsonl
│   ├── long_horizon.jsonl
│   ├── spatial.jsonl
│   └── visual_appearance.jsonl
├── tasks.zip              # Compressed task annotations
└── README.md

Data Fields

Each row in the JSONL files contains:

• id (int): Unique identifier within the split
• task (string): Task path in format {task_type}-{object}-{receptacle}-{scene}/trial_{id}
• repeat_idx (int): Annotation repeat index (0, 1, or 2)
• instruction (string): Natural language instruction for the task
• task_type (string): Category of the task
• trial_id (string): Trial identifier

Task Data

The tasks.zip file contains detailed annotations for each task, including:

• Expert trajectories
• Step-by-step instructions
• Scene information
• Object states

When extracted, the structure is:

tasks/
└── {task_type}-{object}-{receptacle}-{scene}/
    └── trial_{id}/
        └── pp/
            ├── ann_0.json
            ├── ann_1.json
            └── ann_2.json

Usage

Loading with Datasets Library

from datasets import load_dataset

# Load a specific subset
dataset = load_dataset("oscarqjh/EB-Alfred_easi", split="base")

# Load all subsets
dataset = load_dataset("oscarqjh/EB-Alfred_easi")

# Access data
for example in dataset["base"]:
    print(example["instruction"])

Loading Task Annotations

import json
import zipfile
from huggingface_hub import hf_hub_download

# Download the tasks.zip file
zip_path = hf_hub_download(
    repo_id="oscarqjh/EB-Alfred_easi",
    filename="tasks.zip",
    repo_type="dataset"
)

# Extract and read a specific task annotation
task_path = "tasks/pick_and_place_simple-Bowl-None-CoffeeTable-201/trial_T20190907_171639_045329/pp/ann_0.json"
with zipfile.ZipFile(zip_path, 'r') as zipf:
    with zipf.open(task_path) as f:
        annotation = json.load(f)

# Or extract all files to a directory
with zipfile.ZipFile(zip_path, 'r') as zipf:
    zipf.extractall('.')  # Extracts to ./tasks/

Acknowledgements

This dataset is based on the ALFRED dataset and is part of the EmbodiedBench benchmark suite.