Image Segmentation
PyTorch
English
painting-vision-robotics-kit
semantic-segmentation
robotics
edge-ai
construction-ai
autonomous-painting
wall-painting-robot
paint-coverage-estimation
building-facade
drywall
skirting-detection
window-detection
lidar
depth-validation
segformer
deeplabv3
Eval Results (legacy)
Upload safety_gate.py with huggingface_hub
Browse files- safety_gate.py +181 -4
safety_gate.py
CHANGED
|
@@ -5,7 +5,13 @@ A perception stack that always answers is dangerous. This gate combines the
|
|
| 5 |
semantic prediction with the independent depth check and returns a single
|
| 6 |
**approve / reject** decision with the exact reasons. It fails closed: missing
|
| 7 |
evidence (no wall, high uncertainty, no depth confirmation, unaccounted
|
| 8 |
-
protrusions
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
|
| 10 |
Inputs are the JSON outputs already produced by the pipeline:
|
| 11 |
``predict.py --json`` (``--prediction``) and ``depth_validate.py --report``
|
|
@@ -15,6 +21,7 @@ Usage::
|
|
| 15 |
|
| 16 |
python3 safety_gate.py --prediction result.json --depth-report validation.json
|
| 17 |
python3 safety_gate.py --prediction result.json --no-require-depth
|
|
|
|
| 18 |
"""
|
| 19 |
from __future__ import annotations
|
| 20 |
|
|
@@ -29,18 +36,118 @@ DEFAULT_THRESHOLDS = {
|
|
| 29 |
"max_missed_protrusion_fraction": 0.01,
|
| 30 |
"max_missed_recess_fraction": 0.01,
|
| 31 |
"require_depth": True,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
}
|
| 33 |
|
| 34 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
def _number(mapping, key, default=None):
|
| 36 |
value = mapping.get(key, default)
|
| 37 |
return value if isinstance(value, (int, float)) else default
|
| 38 |
|
| 39 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
def assess(prediction, depth=None, thresholds=None):
|
| 41 |
"""Return ``{"decision", "reasons", "checks"}``; ``decision`` is approve/reject."""
|
| 42 |
limits = {**DEFAULT_THRESHOLDS, **(thresholds or {})}
|
| 43 |
-
reasons, checks = [], {}
|
| 44 |
|
| 45 |
wall_pixels = int(_number(prediction, "wall_pixels", 0) or 0)
|
| 46 |
checks["wall_pixels"] = wall_pixels
|
|
@@ -55,6 +162,22 @@ def assess(prediction, depth=None, thresholds=None):
|
|
| 55 |
if prediction.get("inspection_status") == "reobserve_low_confidence":
|
| 56 |
reasons.append("perception requested re-observation (low confidence)")
|
| 57 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
if depth is None:
|
| 59 |
checks["depth_confirmed"] = False
|
| 60 |
if limits["require_depth"]:
|
|
@@ -80,6 +203,36 @@ def assess(prediction, depth=None, thresholds=None):
|
|
| 80 |
if fraction > limits[limit_key]:
|
| 81 |
reasons.append(f"depth shows unaccounted {label} covering {fraction:.1%} of the wall keep-out")
|
| 82 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
return {"decision": "reject" if reasons else "approve", "reasons": reasons, "checks": checks}
|
| 84 |
|
| 85 |
|
|
@@ -91,17 +244,41 @@ def main():
|
|
| 91 |
help="do not reject when depth confirmation is missing (less safe)")
|
| 92 |
parser.add_argument("--max-uncertain-fraction", type=float)
|
| 93 |
parser.add_argument("--min-plane-fraction", type=float)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
args = parser.parse_args()
|
| 95 |
|
| 96 |
-
thresholds = {"require_depth": not args.no_require_depth
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 97 |
if args.max_uncertain_fraction is not None:
|
| 98 |
thresholds["max_uncertain_fraction"] = args.max_uncertain_fraction
|
| 99 |
if args.min_plane_fraction is not None:
|
| 100 |
thresholds["min_plane_fraction"] = args.min_plane_fraction
|
|
|
|
|
|
|
| 101 |
|
| 102 |
prediction = json.loads(args.prediction.read_text(encoding="utf-8"))
|
| 103 |
depth = json.loads(args.depth_report.read_text(encoding="utf-8")) if args.depth_report else None
|
| 104 |
-
result = assess(prediction, depth, thresholds)
|
| 105 |
print(json.dumps(result, indent=2))
|
| 106 |
raise SystemExit(0 if result["decision"] == "approve" else 1)
|
| 107 |
|
|
|
|
| 5 |
semantic prediction with the independent depth check and returns a single
|
| 6 |
**approve / reject** decision with the exact reasons. It fails closed: missing
|
| 7 |
evidence (no wall, high uncertainty, no depth confirmation, unaccounted
|
| 8 |
+
protrusions, possible wallpaper, peeling paint or rust, unassessed surface
|
| 9 |
+
condition) means *do not paint*, never "assume it is fine".
|
| 10 |
+
|
| 11 |
+
``--signoff`` asks the other question: may the robot declare this wall finished
|
| 12 |
+
and move on? It approves only when ``predict.py`` reports ``completion.status ==
|
| 13 |
+
"complete"``, which needs a checkpoint whose completion threshold was validated
|
| 14 |
+
(``train.py`` stores it); otherwise a person inspects the wall.
|
| 15 |
|
| 16 |
Inputs are the JSON outputs already produced by the pipeline:
|
| 17 |
``predict.py --json`` (``--prediction``) and ``depth_validate.py --report``
|
|
|
|
| 21 |
|
| 22 |
python3 safety_gate.py --prediction result.json --depth-report validation.json
|
| 23 |
python3 safety_gate.py --prediction result.json --no-require-depth
|
| 24 |
+
python3 safety_gate.py --prediction after_pass.json --signoff
|
| 25 |
"""
|
| 26 |
from __future__ import annotations
|
| 27 |
|
|
|
|
| 36 |
"max_missed_protrusion_fraction": 0.01,
|
| 37 |
"max_missed_recess_fraction": 0.01,
|
| 38 |
"require_depth": True,
|
| 39 |
+
# Painting over wallpaper is a different job (strip or seal first). A model
|
| 40 |
+
# detection or a pattern/seam cue both stop the robot until someone confirms.
|
| 41 |
+
"allow_wallpaper": False,
|
| 42 |
+
# Peeling paint and rust need preparing (scrape, treat, prime) before a coat.
|
| 43 |
+
"allow_surface_defects": False,
|
| 44 |
+
# A checkpoint without a condition head cannot see paint failure: unknown is not "fine".
|
| 45 |
+
"require_condition": True,
|
| 46 |
+
# The colour-only rust cue scored near chance on 117 reviewed real photos (AUROC
|
| 47 |
+
# ~0.6; it cannot tell a rusted close-up from brick), so on its own it warns
|
| 48 |
+
# rather than stops. The learned condition head is the detector.
|
| 49 |
+
"strict_rust_cue": False,
|
| 50 |
+
# The colour on the wall, and the colour on the tins on site, against the specified one.
|
| 51 |
+
# A wrong colour found after a coat has dried is a repaint, so a mismatch is a reason to
|
| 52 |
+
# stop; a loose job can widen --color-tolerance in predict.py, and this gate honours the
|
| 53 |
+
# verdict predict.py reached. Set max_color_delta_e here to apply the gate's own limit.
|
| 54 |
+
"max_color_delta_e": None,
|
| 55 |
+
# Equipment standing in the paint area. Painting over a customer's tin is a call-back, but
|
| 56 |
+
# the cue detector has known false positives, so by default it warns instead of stopping.
|
| 57 |
+
"strict_objects_in_paint_zone": False,
|
| 58 |
+
# Are the tools the plan needs on site? A missing roller wastes a setup, not a wall.
|
| 59 |
+
"require_tools_on_site": False,
|
| 60 |
}
|
| 61 |
|
| 62 |
|
| 63 |
+
def _condition_reasons(prediction, limits, checks, warnings):
|
| 64 |
+
condition = prediction.get("surface_condition")
|
| 65 |
+
status = (condition or {}).get("status")
|
| 66 |
+
checks["surface_condition_status"] = status
|
| 67 |
+
if limits["allow_surface_defects"]:
|
| 68 |
+
return []
|
| 69 |
+
if status == "defects_detected":
|
| 70 |
+
parts = [f"{name.replace('_', ' ')} on {condition[name]['fraction_of_wall']:.1%} of the wall"
|
| 71 |
+
for name in condition.get("defects", []) if name in condition]
|
| 72 |
+
return [f"surface needs preparation before painting: {', '.join(parts) or 'defects detected'}"]
|
| 73 |
+
weak = (condition or {}).get("weak_damage_signal")
|
| 74 |
+
if weak:
|
| 75 |
+
warnings.append(f"faint paint-failure/rust signal on {weak:.1%} of the wall, below the calibrated "
|
| 76 |
+
"threshold; look before painting")
|
| 77 |
+
if status == "possible_rust":
|
| 78 |
+
score = (condition.get("cue") or {}).get("score")
|
| 79 |
+
message = f"possible rust (colour cue {score}); inspect and treat before painting"
|
| 80 |
+
if limits["strict_rust_cue"]:
|
| 81 |
+
return [message]
|
| 82 |
+
warnings.append(message)
|
| 83 |
+
return []
|
| 84 |
+
if limits["require_condition"] and status in (None, "not_assessed"):
|
| 85 |
+
return ["surface condition not assessed (no condition head in this checkpoint); "
|
| 86 |
+
"inspect for peeling paint and rust"]
|
| 87 |
+
return []
|
| 88 |
+
|
| 89 |
+
|
| 90 |
def _number(mapping, key, default=None):
|
| 91 |
value = mapping.get(key, default)
|
| 92 |
return value if isinstance(value, (int, float)) else default
|
| 93 |
|
| 94 |
|
| 95 |
+
def _color_reasons(prediction, limits, checks, warnings):
|
| 96 |
+
"""The coat (and the tins on site) against the specified colour, when one was given."""
|
| 97 |
+
colour = prediction.get("paint_color") or {}
|
| 98 |
+
objects = prediction.get("site_objects") or {}
|
| 99 |
+
reference = colour.get("reference")
|
| 100 |
+
checks["paint_color_status"] = colour.get("status")
|
| 101 |
+
checks["paint_color_reference"] = (reference or {}).get("verdict")
|
| 102 |
+
reasons = []
|
| 103 |
+
if reference:
|
| 104 |
+
measured = colour.get("dominant") or {}
|
| 105 |
+
delta = _number(reference, "delta_e")
|
| 106 |
+
message = (f"the coat is {measured.get('name')} {measured.get('hex')}, ΔE00 {delta} from the "
|
| 107 |
+
f"specified {(reference.get('reference') or {}).get('name')} "
|
| 108 |
+
f"{(reference.get('reference') or {}).get('hex')}")
|
| 109 |
+
limit = limits["max_color_delta_e"]
|
| 110 |
+
if reference.get("passes"):
|
| 111 |
+
if colour.get("multi_colour"):
|
| 112 |
+
warnings.append(f"{message}: a second colour is also on the wall, check the faces")
|
| 113 |
+
elif limit is not None and delta is not None and delta <= limit:
|
| 114 |
+
warnings.append(message)
|
| 115 |
+
else:
|
| 116 |
+
reasons.append(f"{message}: check the tin before more coat goes on")
|
| 117 |
+
tin = (objects.get("reference_check") or {})
|
| 118 |
+
checks["tins_on_site"] = tin.get("status")
|
| 119 |
+
if tin.get("status") == "tin_differs_from_spec":
|
| 120 |
+
closest = min(tin.get("tins") or [{"delta_e": None}], key=lambda item: item.get("delta_e") or 1e9)
|
| 121 |
+
reasons.append(f"the tins on site do not match the specified colour "
|
| 122 |
+
f"{(tin.get('reference') or {}).get('name')} (closest ΔE00 {closest.get('delta_e')}): "
|
| 123 |
+
"check the product before painting")
|
| 124 |
+
return reasons
|
| 125 |
+
|
| 126 |
+
|
| 127 |
+
def _object_reasons(prediction, limits, checks, warnings):
|
| 128 |
+
"""Equipment found on site: what is in the way, and what the plan needs but cannot see."""
|
| 129 |
+
objects = prediction.get("site_objects") or {}
|
| 130 |
+
checks["site_objects_status"] = objects.get("status")
|
| 131 |
+
checks["equipment_present"] = objects.get("present")
|
| 132 |
+
checks["missing_tools"] = objects.get("missing_tools")
|
| 133 |
+
reasons = []
|
| 134 |
+
clutter = objects.get("in_paint_zone") or []
|
| 135 |
+
if clutter:
|
| 136 |
+
named = ", ".join(sorted({str(item["class"]).replace("_", " ") for item in clutter}))
|
| 137 |
+
message = (f"equipment is standing in the paint area ({named}); it stays out of the plan but "
|
| 138 |
+
"check the way through")
|
| 139 |
+
(reasons if limits["strict_objects_in_paint_zone"] else warnings).append(message)
|
| 140 |
+
missing = objects.get("missing_tools") or []
|
| 141 |
+
if missing:
|
| 142 |
+
message = f"the plan needs {', '.join(missing)} and none is visible on site"
|
| 143 |
+
(reasons if limits["require_tools_on_site"] else warnings).append(message)
|
| 144 |
+
return reasons
|
| 145 |
+
|
| 146 |
+
|
| 147 |
def assess(prediction, depth=None, thresholds=None):
|
| 148 |
"""Return ``{"decision", "reasons", "checks"}``; ``decision`` is approve/reject."""
|
| 149 |
limits = {**DEFAULT_THRESHOLDS, **(thresholds or {})}
|
| 150 |
+
reasons, checks, warnings = [], {}, []
|
| 151 |
|
| 152 |
wall_pixels = int(_number(prediction, "wall_pixels", 0) or 0)
|
| 153 |
checks["wall_pixels"] = wall_pixels
|
|
|
|
| 162 |
if prediction.get("inspection_status") == "reobserve_low_confidence":
|
| 163 |
reasons.append("perception requested re-observation (low confidence)")
|
| 164 |
|
| 165 |
+
wallpaper = prediction.get("wallpaper") or {}
|
| 166 |
+
status = wallpaper.get("status")
|
| 167 |
+
checks["wallpaper_status"] = status
|
| 168 |
+
if not limits["allow_wallpaper"]:
|
| 169 |
+
if status == "wallpaper_detected":
|
| 170 |
+
fraction = _number(wallpaper, "fraction_of_wall", 0.0) or 0.0
|
| 171 |
+
reasons.append(f"wallpaper covers {fraction:.0%} of the wall; strip or seal before painting")
|
| 172 |
+
elif status == "possible_wallpaper":
|
| 173 |
+
cue = _number(wallpaper.get("cue") or {}, "score")
|
| 174 |
+
reasons.append(f"possible wallpaper (pattern/seam cue {cue if cue is not None else '?'}); "
|
| 175 |
+
f"confirm the surface before painting")
|
| 176 |
+
|
| 177 |
+
reasons.extend(_condition_reasons(prediction, limits, checks, warnings))
|
| 178 |
+
reasons.extend(_color_reasons(prediction, limits, checks, warnings))
|
| 179 |
+
reasons.extend(_object_reasons(prediction, limits, checks, warnings))
|
| 180 |
+
|
| 181 |
if depth is None:
|
| 182 |
checks["depth_confirmed"] = False
|
| 183 |
if limits["require_depth"]:
|
|
|
|
| 203 |
if fraction > limits[limit_key]:
|
| 204 |
reasons.append(f"depth shows unaccounted {label} covering {fraction:.1%} of the wall keep-out")
|
| 205 |
|
| 206 |
+
return {"decision": "reject" if reasons else "approve", "reasons": reasons, "warnings": warnings,
|
| 207 |
+
"checks": checks}
|
| 208 |
+
|
| 209 |
+
|
| 210 |
+
def assess_signoff(prediction):
|
| 211 |
+
"""May the robot declare the wall finished? Fails closed like :func:`assess`."""
|
| 212 |
+
reasons, checks = [], {}
|
| 213 |
+
completion = prediction.get("completion") or {}
|
| 214 |
+
status = completion.get("status")
|
| 215 |
+
checks["completion_status"] = status
|
| 216 |
+
checks["confident_painted_fraction"] = _number(completion, "confident_painted_fraction")
|
| 217 |
+
checks["completion_threshold"] = completion.get("completion_threshold")
|
| 218 |
+
if status is None:
|
| 219 |
+
reasons.append("prediction has no completion block (predict.py predates completion checks)")
|
| 220 |
+
elif status == "not_calibrated":
|
| 221 |
+
reasons.append("checkpoint has no validated completion threshold; a person must sign the wall off")
|
| 222 |
+
elif status == "no_wall_detected":
|
| 223 |
+
reasons.append("no wall detected")
|
| 224 |
+
elif status != "complete":
|
| 225 |
+
fraction = checks["confident_painted_fraction"]
|
| 226 |
+
regions = completion.get("touch_up_regions_total", 0)
|
| 227 |
+
reasons.append(f"only {fraction:.0%} of the wall is confidently painted; "
|
| 228 |
+
f"{regions} touch-up region(s) remain" if fraction is not None else f"completion status {status}")
|
| 229 |
+
if prediction.get("inspection_status") == "reobserve_low_confidence":
|
| 230 |
+
reasons.append("perception requested re-observation (low confidence)")
|
| 231 |
+
# A finished coat over rust or failing paint is not finished: it will not last.
|
| 232 |
+
condition = (prediction.get("surface_condition") or {}).get("status")
|
| 233 |
+
checks["surface_condition_status"] = condition
|
| 234 |
+
if condition in ("defects_detected", "possible_rust"):
|
| 235 |
+
reasons.append(f"surface condition {condition.replace('_', ' ')}; the wall cannot be signed off")
|
| 236 |
return {"decision": "reject" if reasons else "approve", "reasons": reasons, "checks": checks}
|
| 237 |
|
| 238 |
|
|
|
|
| 244 |
help="do not reject when depth confirmation is missing (less safe)")
|
| 245 |
parser.add_argument("--max-uncertain-fraction", type=float)
|
| 246 |
parser.add_argument("--min-plane-fraction", type=float)
|
| 247 |
+
parser.add_argument("--allow-surface-defects", action="store_true",
|
| 248 |
+
help="surface already prepared: do not reject on detected paint failure/rust")
|
| 249 |
+
parser.add_argument("--strict-rust-cue", action="store_true",
|
| 250 |
+
help="reject on the colour-only rust cue too (it is near chance on real photos)")
|
| 251 |
+
parser.add_argument("--no-require-condition", action="store_true",
|
| 252 |
+
help="do not reject when the checkpoint cannot assess surface condition (less safe)")
|
| 253 |
+
parser.add_argument("--signoff", action="store_true",
|
| 254 |
+
help="decide whether the wall may be declared finished instead of whether to paint")
|
| 255 |
+
parser.add_argument("--allow-wallpaper", action="store_true",
|
| 256 |
+
help="surface confirmed paintable: do not reject on detected/possible wallpaper")
|
| 257 |
+
parser.add_argument("--max-color-delta-e", type=float,
|
| 258 |
+
help="apply the gate's own colour limit: ΔE00 above it is a mismatch, below it a warning")
|
| 259 |
+
parser.add_argument("--strict-objects-in-paint-zone", action="store_true",
|
| 260 |
+
help="reject on equipment standing in the paint area (the cue detector has known "
|
| 261 |
+
"false positives, so by default this only warns)")
|
| 262 |
+
parser.add_argument("--require-tools-on-site", action="store_true",
|
| 263 |
+
help="reject when a tool the plan needs cannot be seen on site")
|
| 264 |
args = parser.parse_args()
|
| 265 |
|
| 266 |
+
thresholds = {"require_depth": not args.no_require_depth, "allow_wallpaper": args.allow_wallpaper,
|
| 267 |
+
"allow_surface_defects": args.allow_surface_defects,
|
| 268 |
+
"require_condition": not args.no_require_condition,
|
| 269 |
+
"strict_rust_cue": args.strict_rust_cue,
|
| 270 |
+
"strict_objects_in_paint_zone": args.strict_objects_in_paint_zone,
|
| 271 |
+
"require_tools_on_site": args.require_tools_on_site}
|
| 272 |
if args.max_uncertain_fraction is not None:
|
| 273 |
thresholds["max_uncertain_fraction"] = args.max_uncertain_fraction
|
| 274 |
if args.min_plane_fraction is not None:
|
| 275 |
thresholds["min_plane_fraction"] = args.min_plane_fraction
|
| 276 |
+
if args.max_color_delta_e is not None:
|
| 277 |
+
thresholds["max_color_delta_e"] = args.max_color_delta_e
|
| 278 |
|
| 279 |
prediction = json.loads(args.prediction.read_text(encoding="utf-8"))
|
| 280 |
depth = json.loads(args.depth_report.read_text(encoding="utf-8")) if args.depth_report else None
|
| 281 |
+
result = assess_signoff(prediction) if args.signoff else assess(prediction, depth, thresholds)
|
| 282 |
print(json.dumps(result, indent=2))
|
| 283 |
raise SystemExit(0 if result["decision"] == "approve" else 1)
|
| 284 |
|