Datasetpublic

oscarqjh/REVERIE-CE_easi

REVERIE-CE (EASI) Pre-converted REVERIE benchmark episodes for continuous environments (Habitat-Sim), formatted for the EASI evaluation framework. Source Episodes are derived from Dynam3D's pre-converted REVERIE-CE data, which converts the original REVERIE discrete navigation graph episodes into continuous Habitat-Sim environments using Matterport3D scenes. Navigation-only: This dataset evaluates navigation to the described location. Object grounding is not… See the full description on the dataset page: https://huggingface.co/datasets/oscarqjh/REVERIE-CE_easi.

sourceHugging Faceotherupdated 7mo agoView on Hugging Face

0likes31downloads

Dataset Card

REVERIE-CE (EASI)

Pre-converted REVERIE benchmark episodes for continuous environments (Habitat-Sim), formatted for the EASI evaluation framework.

Source

Episodes are derived from Dynam3D's pre-converted REVERIE-CE data, which converts the original REVERIE discrete navigation graph episodes into continuous Habitat-Sim environments using Matterport3D scenes.

Navigation-only: This dataset evaluates navigation to the described location. Object grounding is not included.

Dataset Structure

data/
├── val_unseen.jsonl       # 3,189 episodes (unseen environments, 26 scenes)
└── test.jsonl             # 6,292 episodes
mp3d_scenes.zip            # Matterport3D .glb scene files (26 scenes, 1.5 GB)

Episode Format (JSONL)

Each line is a JSON object:

json

{
  "episode_id": "1",
  "scene_id": "zsNo4HB9uLZ",
  "instruction": "Enter the dining room and pick up the objects on the table",
  "start_position": [1.395, 0.299, -4.876],
  "start_rotation": [0.0, 0.367, 0.0, 0.930],
  "goal_position": [14.052, 0.312, -1.990],
  "geodesic_distance": 13.37,
  "gt_locations": [[1.395, 0.299, -4.876], ..., [14.052, 0.312, -1.990]]
}

Field	Description
`episode_id`	Unique episode identifier
`scene_id`	Matterport3D scan name
`instruction`	High-level natural language instruction (REVERIE style)
`start_position`	Agent start position in Habitat coordinates `[x, y, z]`
`start_rotation`	Start rotation quaternion `[x, y, z, w]`
`goal_position`	Target position (navigation goal)
`geodesic_distance`	Shortest walkable path distance from start to goal (meters)
`gt_locations`	Ground truth waypoints from start to goal

Matterport3D Scenes

mp3d_scenes.zip contains .glb scene files for Habitat-Sim. These are the same Matterport3D scenes used by VLN-CE R2R. After extraction:

mp3d/
├── {scan_id}/
│   └── {scan_id}.glb
├── ...

Note: Matterport3D scenes require agreeing to the Matterport3D Terms of Use.

Metrics

Metric	Description
SR (Success Rate)	1.0 if agent stops within 3m of goal
SPL	Success weighted by path efficiency
NE (Navigation Error)	Geodesic distance to goal at stop
Oracle SR	Best geodesic distance achieved during episode
NDTW	Normalized Dynamic Time Warping
SDTW	Success-weighted DTW

Usage with EASI

bash

easi task download reverie_ce_val_unseen
easi start reverie_ce_val_unseen --agent react --backend openai --model gpt-4o

References

—REVERIE — Original discrete REVERIE benchmark (Qi et al., CVPR 2020)
—Dynam3D — Discrete-to-CE conversion (Wang et al., NeurIPS 2025)
—VLN-CE — Vision-Language Navigation in Continuous Environments (Krantz et al., 2020)

License

Episode data follows the licenses of the original REVERIE and Dynam3D datasets. Matterport3D scenes are subject to the Matterport3D Terms of Use.