Generate additional search tasks

Select the original disaster layout with --scene, such as fire/urban/level_1. The generator loads that scene’s people-free damaged geometry and its own verified placement candidates. It samples survivor positions and poses on that layout while keeping buildings, debris and other damaged geometry fixed.

./dw survivors generate --scene fire/urban/level_1 \
  --seed 7 --count 20 --output data/disasterworld/generated/fire-urban-l1-task-7

Then run the generated task on the same scene:

./dw run --scene fire/urban/level_1 \
  --generated-task generated/fire-urban-l1-task-7/generated_task.json \
  --robots 4 --baseline frontier

--scene selects scenes/fire/urban/level_1/manifest.json in this example. Its geometry_file and placement_candidates_file identify the original geometry and the candidates validated for that exact layout. The generator checks the geometry checksum before sampling, and the task records its source scene and geometry checksum. A run rejects a task if its scene or geometry differs.

Generated tasks live inside dataset.root so USD references work in the simulator container. generated_task is a dataset-relative manifest path. Published variants remain separate; generated tasks are marked synthetic in evaluation metadata. Candidate sampling is deterministic for a given seed and occurs without replacement; a location can have several verified posed assets.

The publisher’s placement_candidates.json must contain the SHA-256 of geometry_file and candidates like:

{
  "geometry_sha256": "SHA256_OF_PEOPLE_FREE_ROOT_LAYER",
  "candidates": [{
    "id": "candidate_01",
    "position": [10,20,0.5],
    "verified_static_collision": true,
    "poses": [{
      "asset_file": "survivor_assets/lying.usda",
      "pose_label": "lying",
      "yaw_degrees": 45,
      "verified_static_collision": true
    }]
  }]
}

Posed assets must have a valid default prim and an origin compatible with the candidate translation. Verification applies to the full posed body at that candidate. The generator trusts publisher verification metadata; it does not run a physics clearance check. Candidate locations must also avoid mutually overlapping bodies when sampled together. The current SHA checks the root geometry layer, so the publisher must freeze referenced dependencies as well.

The generator writes a USD scene that references the selected geometry and adds the sampled survivors, along with ground truth and a task manifest.