Configuration¶

The data generator uses YAML config files with a config hierarchy and CLI overrides. All paths are relative to $NAVARENA_DATA_DIR.

Config Hierarchy¶

Main config - User-provided full config (e.g. examples/pointnav_example.yaml)
defaults/env/{env_type}.yaml - Environment defaults (e.g. gs.yaml)
defaults/task/{task_type}.yaml - Task defaults (e.g. pointnav.yaml)
defaults/planner.yaml - Trajectory planner defaults

User config wins; missing fields are merged from the above defaults.

Config Structure¶

configs/
├── defaults/
│   ├── env/
│   │   └── gs.yaml                 # Only gs (3DGS) is implemented; habitat/isaac are placeholders
│   ├── task/
│   │   ├── pointnav.yaml
│   │   ├── gridtraj.yaml
│   │   ├── imagenav.yaml
│   │   ├── objectnav.yaml
│   │   └── vln.yaml
│   └── planner.yaml
└── examples/
    ├── pointnav_example.yaml
    ├── gridtraj_example.yaml
    ├── imagenav_example.yaml
    ├── objectnav_example.yaml
    ├── vln_zh_example.yaml
    └── ...

Example Configs¶

PointNav¶

env_type: gs
scene_path: sage-3d/00666b7a

env_config:
  z_coordinate: 0.0
  robot_radius: 0.4
  max_linear_velocity: 0.5
  max_angular_velocity: 1.0

task_type: pointnav
num_episodes: 1000
split: train

task_config:
  start_constraints:
    sampler: grid
    grid_spacing: 0.5
    max_start_points: null

  goal_constraints:
    sampler: grid
    grid_spacing: 0.5
    require_navigable_path: true

  trajectory_constraints:
    min_geodesic_distance: 0.5
    max_geodesic_distance: 30.0
    require_gt_path: true

dataset_name: navarena_dataset_v1

VLN (Chinese Instructions)¶

env_type: gs
scene_path: x2robot/17dc3367

env_config:
  z_coordinate: 0.0
  robot_radius: 0.4

task_type: vln
num_episodes: 100
split: train

task_config:
  start_constraints:
    sampler: grid
    grid_spacing: 0.5
  goal_constraints:
    sampler: grid
    grid_spacing: 0.5
  trajectory_constraints:
    min_geodesic_distance: 5.0
    max_geodesic_distance: 12.0

  instruction_type: simple_direction   # simple_direction / path_based / object_goal
  language: zh-CN
  num_instructions_per_episode: 1

dataset_name: navarena_vln_zh

Parameter Reference¶

Top-Level¶

Parameter	Description	Default
`env_type`	Environment type: gs / habitat / isaac	`gs`
`scene_path`	Scene path, e.g. `x2robot/17dc3367`	Required
`task_type`	Task type: pointnav / imagenav / objectnav / vln	`pointnav`
`num_episodes`	Number of episodes to generate	`100`
`split`	Data split: train / val_seen / val_unseen / test	`train`
`dataset_name`	Dataset name, output under `datasets/{dataset_name}/`	`navarena_dataset_v1`

env_config¶

Parameter	Description	Default
`z_coordinate`	Fixed height	`0.0`
`robot_radius`	Collision radius (meters)	`0.4`
`max_linear_velocity`	Max linear velocity (m/s)	`0.5`
`max_angular_velocity`	Max angular velocity (rad/s)	`1.0`
`max_linear_accel`	Max linear acceleration	`0.3`
`max_angular_accel`	Max angular acceleration	`2.0`
`max_jerk`	Max jerk	`0.5`
`max_lateral_accel`	Max lateral acceleration	`0.3`
`dt`	Time step (seconds)	`0.333`
`goal_tolerance`	Goal tolerance (meters)	`0.3`

task_config¶

start_constraints¶

Parameter	Description	Default
`sampler`	Sampler: grid	`grid`
`grid_spacing`	Grid spacing (meters)	`0.5`
`max_start_points`	Max start points, null = unlimited	`null`
`rotation_num`	Orientations per point	`6`

goal_constraints¶

Parameter	Description	Default
`sampler`	Sampler: grid	`grid`
`grid_spacing`	Grid spacing (meters)	`0.5`
`rotation_num`	Orientations per goal	`6`
`require_navigable_path`	Goal must have navigable path from start	`true`

trajectory_constraints¶

Parameter	Description	Default
`min_geodesic_distance`	Min geodesic distance (meters)	`0.5`
`max_geodesic_distance`	Max geodesic distance (meters)	`30.0`
`require_gt_path`	Require valid GT path	`true`

VLN-specific¶

Parameter	Description	Default
`instruction_type`	simple_direction / path_based / object_goal	Required
`language`	zh-CN / en-US	-
`num_instructions_per_episode`	Instructions per episode	`1`

planner_config (optional)¶

Configure via task_config.planner_config or defaults/planner.yaml; passed to the trajectory planner.

Multi-File Merge¶

Use GeneratorConfig.from_files() or --env-config / --task-config to merge fragments:

python scripts/generate_data.py --config configs/examples/pointnav_example.yaml \
    --env-config my_env.yaml --task-config my_task.yaml

CLI Overrides¶

# Config file
python scripts/generate_data.py --config configs/examples/pointnav_example.yaml

# CLI only
python scripts/generate_data.py --env gs --task pointnav \
    --scene x2robot/17dc3367 --num-episodes 100

# Parallel
python scripts/generate_data.py --config configs/examples/pointnav_example.yaml \
    --parallel --num-workers 4 --batch-size 20

# Resume from checkpoint
python scripts/generate_data.py --config configs/examples/pointnav_example.yaml --resume

# Append mode
python scripts/generate_data.py --config configs/examples/pointnav_example.yaml --append

See also: Batch Processing · Asset Preprocessing