miaodx/roboclaws/raw-fpv-visual-labeler

RAW-FPV Visual Labeler

Use this skill only for perception-only RAW-FPV labeling probes. It consumes public robot FPV frame evidence and emits structured visual labels for review and offline scoring. It does not call cleanup tools, does not create observed_* handles, and does not authorize navigation, pick, place, or done.

Inputs

Use 3-6 neighboring RAW-FPV frames from the same waypoint, sweep segment, or source observation neighborhood when available. Each frame may include only:

• public frame id;
• image artifact;
• public waypoint or room context already visible to the cleanup agent;
• optional public runtime-map planning hints marked non-executable.

Never include private labels, generated hidden target ids, acceptable destination truth, executable observed-object handles, detector candidates, or camera-label producer candidates.

Output

Return strict JSON:

{
  "schema": "raw_fpv_visual_labeler_response_v1",
  "labels": [
    {
      "evidence_frame_id": "run/raw_fpv_001",
      "category": "mug",
      "category_family": "dish",
      "coarse_region": "middle_right",
      "confidence": 0.82,
      "is_cleanup_relevant": true,
      "bbox": [0.62, 0.5, 0.12, 0.15],
      "surface_hint": "table",
      "reason_not_actionable": ""
    }
  ]
}

Required per label:

• evidence_frame_id
• category
• category_family
• coarse_region
• confidence
• is_cleanup_relevant

Optional per label:

• bbox
• surface_hint
• reason_not_actionable

Allowed category families are food, dish, book, linen, toy, and electronics. Fixtures and surfaces such as tables, beds, counters, shelves, sinks, cabinets, and floors may be mentioned only as surface_hint or as is_cleanup_relevant=false; they are not object hits.

Boundary

These labels are perception evidence. The cleanup agent must not consume them as executable handles in camera-raw-fpv. A later assisted RAW-FPV or camera-grounded-labels producer decision would need its own contract change.