oscarqjh/manipulathor_easi

ManipulaTHOR APND Dataset

The Arm Point Navigation Dataset (APND) from ManipulaTHOR, converted to HuggingFace format for use with the EASI evaluation framework.

Dataset Description

Each episode defines an object rearrangement task: an agent must pick up a specified object from a source location and move it to a target location within an AI2-THOR kitchen scene. The dataset provides deterministic task configurations with pre-computed source/target positions, agent poses, and arm starting poses.

Splits (matching ManipulaTHOR paper, Table 1)

Split	Scenes	Objects	Episodes
train	19 train scenes	6 seen (Apple, Bread, ...)	6840
val_seen_obj	5 val scenes	6 seen (Apple, Bread, ...)	1800
val_novel_obj	5 val scenes	6 novel (Potato, Pan, ...)	1800
test_seen_obj	5 test scenes	6 seen (Apple, Bread, ...)	1800
test_novel_obj	5 test scenes	6 novel (Potato, Pan, ...)	1800
seen_scenes_novel_obj	19 train scenes	6 novel (Potato, Pan, ...)	6840

Total: 20880 episodes across 30 AI2-THOR kitchen scenes (scene 28 excluded).

Data Fields

Each row in the JSONL files contains:

Field	Type	Description
id	int	Sequential episode ID within the split
scene	string	AI2-THOR scene name (e.g., FloorPlan1_physics)
object_type	string	Object category (e.g., Apple, Mug)
object_id	string	AI2-THOR object instance ID
source_position	JSON string	Source object 3D position (x, y, z)
target_position	JSON string	Target object 3D position (x, y, z)
source_countertop_id	string	Receptacle ID at source location
target_countertop_id	string	Receptacle ID at target location
source_agent_pose	JSON string	Agent pose when viewing source
target_agent_pose	JSON string	Agent pose when viewing target
source_visibility	bool	Whether object is visible from source agent pose
target_visibility	bool	Whether object is visible from target agent pose
initial_agent_pose	JSON string	Starting agent pose for the episode
arm_starting_pose	JSON string	Starting arm configuration for the scene
instruction	string	Task instruction

Dataset Structure

.
├── data/
│   ├── train.jsonl
│   ├── val_seen_obj.jsonl
│   ├── val_novel_obj.jsonl
│   ├── test_seen_obj.jsonl
│   ├── test_novel_obj.jsonl
│   └── seen_scenes_novel_obj.jsonl
├── apnd-dataset.zip          # Original dataset archive
└── README.md

Usage

Loading with Datasets Library

from datasets import load_dataset

# Load the main benchmark split (novel objects, unseen scenes)
dataset = load_dataset("oscarqjh/manipulathor_easi", split="test_novel_obj")

# Access data
for example in dataset:
    print(example["scene"], example["object_type"])
    print(example["instruction"])

Using with EASI

easi start manipulathor_test_novel_obj --agent react --backend openai --model gpt-4o

Acknowledgements

This dataset is derived from the ManipulaTHOR Arm Point Navigation Dataset (APND) and uses AI2-THOR as the simulation environment.

Citation:

@inproceedings{ehsani2021manipulathor,
  title={ManipulaTHOR: A Framework for Visual Object Manipulation},
  author={Ehsani, Kiana and Han, Winson and Herrasti, Alvaro and VanderBilt, Eli and Weihs, Luca and Kolve, Eric and Kembhavi, Aniruddha and Mottaghi, Roozbeh},
  booktitle={CVPR},
  year={2021}
}