Configure and run

Edit configs/default.yaml or create an override YAML. Paths are relative to the configuration file, not the shell’s current directory. Unknown keys are errors.

scene: fire/suburban/level_1
survivor_variant: 1
robots: 4
baseline: frontier
mission_seconds: 600
search:
  region: affected_area
  assignment: partitioned
inference:
  vlm_url: null
  vlm_model: null
  gpu: 0
simulation:
  gpu: 0

scene selects disaster/locality/damage level; survivor_variant selects a published survivor layout. robots uses the first N verified spawn points from the scene manifest. Missions share a start time and a simulated-time budget after takeoff. Policy startup and mapper initialization currently count within the mission budget; HTTP inference services are warmed beforehand. Fixed navigation limits are configured in MIGHTY; the runtime uses the upstream navigation interface.

Search areas

region: affected_area uses the bundle’s disaster boundary; whole_map uses the map boundary. custom uses custom_polygon. These polygons must be convex and given in world XY meters. assignment: partitioned splits the polygon into deterministic equal-area strips. shared gives every robot the entire polygon. The generated assignments.svg previews the allocation.

./dw run --robots 4 --baseline frontier
./dw run --robots 4 --baseline conavgpt

The CoNavGPT command selects whole_map and shared automatically. It starts one team planner on a bridged DDS domain, rather than a separate cooperative planner per robot. Frontier, Lawnmower and VLFM use per-robot policies. RAVEN uses its RayFronts adapter; consult RayFronts’ own docs for mapper usage.

Inference

YOLO/SAM and BLIP run locally through asm_offboard when needed by the selected baseline. Local services share inference.gpu.

Leave inference.vlm_url as null, none, or blank to run the bundled VLM locally. To use an existing OpenAI-compatible vision API, set vlm_url to its API base address and vlm_model to the provider’s model name. The API key comes from the environment variable named by vlm_api_key_env, default DW_VLM_API_KEY. The local VLM service is skipped when a URL is supplied.

See VLM API configuration for a complete example and the supported request format.

Artifacts and shutdown

run validates configuration and generates all mission files before starting containers. For optional inspection, ./dw config check validates settings without running a mission, and ./dw prepare --robots 4 --baseline frontier --output .local/preview writes mission files and an assignment preview. Neither command is required before run.

./dw run --dry-run resolves the launch without starting containers. A completed run saves configuration, version pins, per-robot reports/trajectories, policy logs and container logs under results_root. Policies receive public bounds and observations; survivor ground truth is kept for the host evaluator. ./dw down stops the current DisasterWorld project. Use distinct runtime.project and network_subnet values for simultaneous installations.