oscarqjh/EB-Habitat_easi

EB-Habitat Dataset for EASI

This dataset is a preprocessed version of the EB-Habitat benchmark from EmbodiedBench, converted to HuggingFace-compatible format for use with the EASI evaluation suite.

Overview

EB-Habitat is a language-guided rearrangement benchmark using the Habitat simulator with ReplicaCAD scenes. Agents must follow natural language instructions to navigate to objects, pick them up, and place them at target locations in realistic 3D environments.

Simulator: Habitat (with ReplicaCAD scenes, YCB objects, Fetch robot) Actions: 70 discrete actions (navigation, manipulation, interaction) Max steps: 30 per episode Episodes: 300 total (50 per split)

Splits

Split	Episodes	Description
base	50	Standard rearrangement tasks
common_sense	50	Tasks requiring common sense reasoning
complex_instruction	50	Tasks with complex natural language instructions
spatial_relationship	50	Tasks involving spatial relationships
visual_appearance	50	Tasks requiring visual appearance recognition
long_horizon	50	Tasks requiring many sequential steps

Data Format

Each episode is a JSON object in the JSONL files with these fields:

Field	Type	Description
id	int	Sequential index (0-49) within split
episode_id	string	Original episode ID
instruction	string	Natural language task instruction
instruct_id	string	Instruction template ID
scene_id	string	Habitat scene config path
scene_dataset_config	string	Scene dataset config path
sampled_entities	object	Source/target receptacle and object names
goal_preds	object	PDDL goal expression tree
start_preds	list[str]	Initial state predicates
subgoals	list[list[str]]	Ordered subgoal predicates
start_position	list[float]	Agent start position [x, y, z]
start_rotation	list[float]	Agent start rotation quaternion [x, y, z, w]
rigid_objs	list[object]	Rigid objects with names and 4x4 transforms
ao_states	object	Articulated object states {ao_id: {joint_id: value}}
markers	list[object]	Marker definitions with name, type, and params
name_to_receptacle	object	Mapping from object name to receptacle name
additional_obj_config_paths	list[str]	Extra object config paths
info	object	Episode metadata

Simulator Assets

The simulator_data.zip file contains the simulator assets needed to run episodes:

• replica_cad/ — ReplicaCAD scenes, objects, navmeshes, stages, URDF configs
• objects/ycb/ — YCB object meshes and configs
• robots/hab_fetch/ — Fetch robot model
• datasets/ — Rearrangement dataset configs
• default.physics_config.json — Physics simulation config

Usage with EASI

from datasets import load_dataset

ds = load_dataset("oscarqjh/EB-Habitat_easi", split="base")
print(ds[0]["instruction"])
# "Find a toy airplane and move it to the right counter."

Source

Converted from EmbodiedBench EB-Habitat benchmark.

Citation:

@article{embodiedbench,
  title={EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents},
  author={Rui Yang and Hanyang Chen and Junyu Zhang and Mark Coates},
  year={2025}
}