Datasetpublic

oscarqjh/HAZARD_easi

HAZARD for EASI Preprocessed HAZARD benchmark (ICLR 2024) for the EASI evaluation suite. HAZARD evaluates embodied agents in dynamically changing disaster environments built on the ThreeDWorld (TDW) simulator. Agents must rescue target objects from fire, flood, or wind hazards. Scenarios Scenario (Subset) Train Test Total Max Steps Hazard fire 76 25 101 1500 Spreading flames destroy objects flood 76 25 101 1500 Rising water damages non-waterproof… See the full description on the dataset page: https://huggingface.co/datasets/oscarqjh/HAZARD_easi.

sourceHugging Faceupdated 7mo agoView on Hugging Face

0likes26downloads

Dataset Card

HAZARD for EASI

Preprocessed HAZARD benchmark (ICLR 2024) for the EASI evaluation suite.

HAZARD evaluates embodied agents in dynamically changing disaster environments built on the ThreeDWorld (TDW) simulator. Agents must rescue target objects from fire, flood, or wind hazards.

Scenarios

Scenario (Subset)	Train	Test	Total	Max Steps	Hazard
`fire`	76	25	101	1500	Spreading flames destroy objects
`flood`	76	25	101	1500	Rising water damages non-waterproof objects
`wind`	80	20	100	3000	Wind blows objects away
Total	232	70	302

Usage

python

from datasets import load_dataset

# Load fire scenario, train split
ds = load_dataset("oscarqjh/HAZARD_easi", "fire", split="train")

# Load wind scenario, test split
ds = load_dataset("oscarqjh/HAZARD_easi", "wind", split="test")

Schema

Field	Type	Description
`id`	int	Sequential index within split
`episode_id`	string	Episode directory name (e.g., `mm_craftroom_2a-1`)
`task`	string	Scenario: `fire`, `flood`, or `wind`
`instruction`	string	Natural language task description
`scene_name`	string	TDW scene name (e.g., `mm_craftroom_2a`)
`agent_position`	list[float]	Agent spawn position `[x, y, z]`
`target_categories`	list[string]	Target object category names
`target_object_ids`	list[int]	Target object IDs in the scene
`hazard_params`	string (JSON)	Scenario-specific hazard parameters
`containers`	list[int]	Safe storage object IDs (wind only, empty for fire/flood)
`objects`	string (JSON)	All scene objects: `[{id, name, category, position}]`
`max_steps`	int	Maximum episode steps (1500 or 3000)
`source_dir`	string	Relative path to episode dir in `simulator_data.zip`

Hazard Parameters

Fire: {"fire_positions": [[x, y, z], ...]}

Flood: {"source": [[x, y, z]], "direction": [[dx, dy, dz]], "speed": [float], "flood_source_from": "x_max|x_min|z_max|z_min"}

Wind: {"wind_direction": [x, y, z], "wind_resistance": {"obj_id": 0|1|2}}

Simulator Data

The simulator_data.zip file contains the raw episode data needed by the TDW bridge:

—room_setup_fire/ — Episode directories with info.json, log.txt, flood.json
—room_setup_wind/ — Episode directories with info.json, log.txt
—meta_data/ — Object value and temperature metadata

The log.txt files contain TDW command sequences that define scene layouts and object placements.

Source

—Paper: HAZARD Challenge: Embodied Decision Making in Dynamically Changing Environments (ICLR 2024)
—Code: UMass-Foundation-Model/HAZARD
—Simulator: ThreeDWorld (TDW)

License

Please refer to the HAZARD repository for license information.