oscarqjh/VLN-CE-R2R_easi

VLN-CE R2R Dataset for EASI

Vision-and-Language Navigation in Continuous Environments (VLN-CE) Room-to-Room (R2R) benchmark, repackaged for the EASI evaluation framework.

Task

An agent receives a natural language navigation instruction and must navigate through a Matterport3D indoor environment to reach a goal location. The agent uses discrete actions: STOP, MOVE_FORWARD (0.25m), TURN_LEFT (15 deg), TURN_RIGHT (15 deg).

Success is measured when the agent stops within 3.0m of the goal position.

Splits

Split	Episodes
val_seen	778
val_unseen	1839
test	3408
Total	6025

Note: The test split does not include ground truth trajectories. SR and SPL can be computed locally, but NDTW/SDTW require ground truth and are only available for val_seen and val_unseen.

Data Fields

Each episode row contains:

Field	Type	Description
id	int	Sequential index within split
episode_id	string	Original VLN-CE episode ID
scene_id	string	Matterport3D scene name
instruction	string	Natural language navigation instruction
start_position	list[float]	Agent start [x, y, z] in meters
start_rotation	list[float]	Agent start quaternion [x, y, z, w]
goal_position	list[float]	Goal [x, y, z] in meters
goal_radius	float	Success radius (3.0m)
reference_path	string (JSON)	Waypoint path from start to goal
geodesic_distance	float	Shortest path distance in meters
gt_locations	string (JSON) or null	Dense ground truth waypoints (null for test)
gt_actions	string (JSON) or null	Ground truth action sequence (null for test)

Scene Data

mp3d_scenes.zip contains Matterport3D .glb and .navmesh files for the 85 scenes used across all evaluation splits. After extraction:

mp3d/
  {scene_id}/
    {scene_id}.glb        # 3D mesh for Habitat-Sim rendering
    {scene_id}.navmesh    # Navigation mesh for pathfinding

Only .glb (rendering) and .navmesh (pathfinding) are included. Semantic files (.house, _semantic.ply) are not needed for VLN-CE R2R since goals are specified as coordinates, not semantic object targets.

Metrics

Metric	Description
SR	Success Rate (agent within 3.0m of goal)
SPL	Success weighted by Path Length
NDTW	Normalized Dynamic Time Warping (path shape similarity)
SDTW	Success-weighted NDTW
Oracle SR	Best achievable success along trajectory
NE	Navigation Error (distance to goal at episode end)
Path Length	Total distance traveled

Original Sources

• VLN-CE (Krantz et al.)
• R2R Dataset (Anderson et al.)
• Matterport3D (Chang et al.)

Citation

@inproceedings{krantz2020vlnce,
  title={Beyond the Nav-Graph: Vision-and-Language Navigation in Continuous Environments},
  author={Jacob Krantz and Erik Wijmans and Arjun Majumdar and Dhruv Batra and Stefan Lee},
  booktitle={ECCV},
  year={2020}
}