Data Generator Overview¶

The Data Generator (navarena-gen) is a multi-task Vision-Language Navigation (VLN) data generation framework for PointNav, ImageNav, ObjectNav, and VLN tasks. It generates Episode data from 3D Gaussian Splatting scenes using grid sampling, path planning, and instruction generation.

Core Features¶

• Multi-Task Support - PointNav, ImageNav, ObjectNav, VLN
• Multi-Environment Support - 3D Gaussian Splatting (implemented), Habitat / Isaac (placeholder)
• V1 Asset Format - Reads unified format under $NAVARENA_DATA_DIR/assets/ (manifest.json, nav_map.pgm, etc.)
• Flexible Instruction Generation - Strategy pattern: simple_direction, path_based, object_goal; Chinese and English
• Parallel Processing - Multi-worker episode generation
• Crash Recovery - Lightweight checkpoint, resume from last progress after crash
• Web Viewer - Interactive data browsing

Architecture¶

flowchart TB
    subgraph Input[Input]
        Scene[V1 Asset Scene]
        Config[YAML Config]
    end

    subgraph Core[Core Flow]
        Env[Env Init]
        Gen[Episode Gen]
        Inst[Instruction Gen]
        Write[Data Write]
        Render[Render Optional]

        Env --> Gen --> Inst --> Write --> Render
    end

    subgraph Envs[Environment Types]
        GS[3D GS]
        Habitat[Habitat]
        Isaac[Isaac Sim]
    end

    subgraph Tasks[Task Types]
        PN[PointNav]
        IN[ImageNav]
        ON[ObjectNav]
        VLN[VLN]
    end

    Input --> Core
    Envs --> Env
    Tasks --> Gen

Data Generation Flow¶

1. Environment Init - Load manifest.json, nav_map.pgm, nav_map.yaml from V1 assets; optionally labels.json, nav_mask.png
2. Episode Generation - Task-specific generators: grid start sampling + goal sampling + GT trajectory planning (A*)
3. Instruction Generation - VLN uses Strategy pattern: simple_direction / path_based / object_goal; zh-CN, en-US
4. Data Write - Stream to Parquet chunks (meta/episodes.parquet + data/chunk-NNN/trajectories.parquet), with checkpoint for crash recovery
5. Rendering (Optional) - Goal images (ImageNav), trajectory videos

Output Structure¶

All paths are relative to $NAVARENA_DATA_DIR:

datasets/{dataset_name}/
├── dataset_meta.json                   # Dataset-level metadata
└── {scene_path}/                       # e.g. x2robot/17dc3367
    ├── scene_meta.json                 # Scene metadata
    └── {task_type}/                    # pointnav/imagenav/objectnav/vln
        ├── meta/
        │   ├── info.json               # Task-level metadata
        │   └── episodes.parquet        # Episode index (Parquet v1.0.0)
        ├── data/
        │   └── chunk-NNN/              # Chunked storage (default 1000 episodes/chunk)
        │       ├── trajectories.parquet
        │       └── episodes.parquet
        ├── goal_images/                # ImageNav goal images (optional)
        └── rendered_videos/            # Rendered videos (optional)

Episode Format (Logical Structure)¶

Episodes stored in Parquet can be understood at the application layer as:

{
  "episode_id": "train_000001",
  "scene_path": "x2robot/17dc3367",
  "task_type": "vln",
  "start_state": {
    "position": [1.5, -0.8, 0.0],
    "rotation": [0.0, 0.0, 0.0, 1.0]
  },
  "goals": [...],
  "instructions": [
    {"instruction_text": "Walk about 8 meters to the northeast", "language": "en-US"}
  ],
  "gt_path": {
    "stats": {"geodesic_distance": 8.0, "num_steps": 25}
  }
}

Actual storage: meta/episodes.parquet and data/chunk-NNN/trajectories.parquet, linked via chunk_index. See Navigation Training Data Format for details.

Use Cases¶

1. Training Data - Generate Episodes for VLN models
2. Evaluation Data - Generate PointNav, ImageNav, ObjectNav, VLN evaluation sets
3. Trajectory Rendering - Visualize GT trajectories in videos

Dependencies¶

The data generator depends on asset preprocessing output in V1 format. Use navarena-forge first to preprocess scenes.

See also: Pipeline Stages · Configuration · Batch Processing