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)
File size: 16,555 Bytes
4eecd03 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 | """The nozzle's path through time: what the drone flies, not only where it was planned to go.
``drone_areas.simulate_deposit`` lays paint down from the *geometry* of the passes: every stroke becomes one Gaussian
band, whatever its length, speed or standoff. That is enough to compare plans, and it cannot see the defects a real pass
has, because those are all in the timing:
* a controller that slows down inside a pass leaves a heavy band where it lingered;
* a gun that opens late leaves the first few centimetres thin (``application.trigger_offsets``);
* a nozzle that drifts closer to the wall widens the fan and thins the film;
* a flight log records all three, and none of them survives being reduced to a polyline.
This module turns either a planned sortie or a flight log into a ``Timeline``: samples of time, wall position, standoff,
speed and valve flow. ``drone_areas.deposit_from_timeline`` integrates it, so the film (and the twin's belief) carries
the timing and not only the drawing.
Assumptions, stated: the valve opens and closes linearly over the tip's latency; the drone holds its planned speed
inside a pass unless ``decel_mm`` says the controller slows there; and a planned pass is flown along the planner's
straight line. A log has none of those assumptions: its samples are what the aircraft reported.
Wall frame: x mm along the facade from its A end, z mm up from the ground, standoff the nozzle's distance from the
wall. A planned pass is sampled at 10 mm along its length; a log keeps its own samples.
"""
from __future__ import annotations
import math
from dataclasses import dataclass, field
import numpy as np
import application
DS_MM = 10.0 # arc-length sampling of a planned pass
@dataclass
class Timeline:
"""Samples of a sortie, all arrays the same length and in time order."""
t: np.ndarray # s since the sortie started
x_mm: np.ndarray # along the facade from A
z_mm: np.ndarray # up from the ground
standoff_mm: np.ndarray # nozzle distance from the wall
speed_mm_s: np.ndarray # along-track speed
flow: np.ndarray # 0-1 valve flow (pump PWM normalised); 0 = not spraying
pass_id: np.ndarray # int, -1 in transit or line-up
phase: np.ndarray # "lineup" | "spray" | "transit"
sortie: int = 1
source: str = "planned"
spray: dict = field(default_factory=dict)
mm_per_px: float = 1.0
shape: tuple = (0, 0)
ramps: bool = True # planned: the valve ramps between samples; log: its state holds until the next sample
reaction_n: float = 0.0 # the spray's push back on the nozzle, N (application.spray_reaction)
sag_mm: float = 0.0 # the steady standoff the reaction costs at the modelled hold stiffness
standoff_bias_mm: float = 0.0 # what the controller was asked to bias its setpoint by to cancel that sag
def __len__(self):
return int(len(self.t))
def flow_mid(self):
"""Flow over each interval: the ramp's midpoint for a plan, the held state for a log (which samples a switch)."""
if len(self) < 2:
return np.zeros(0)
if self.ramps:
return 0.5 * (self.flow[:-1] + self.flow[1:])
return self.flow[:-1].astype(float)
def intervals(self):
"""dt plus the mid-interval speed and flow: a gap *between* passes is not charged to the sample after it."""
if len(self) < 2:
return np.zeros(0), np.zeros(0), np.zeros(0)
return np.diff(self.t), 0.5 * (self.speed_mm_s[:-1] + self.speed_mm_s[1:]), self.flow_mid()
def stats(self):
"""What the sortie did: time, spray time, paint pumped, and the path flown (all of it, and while spraying)."""
if not len(self):
return {"duration_s": 0.0, "spray_s": 0.0, "pump_on_s": 0.0, "paint_l": 0.0, "path_m": 0.0,
"spray_path_m": 0.0, "reaction_n": 0.0, "standoff_sag_mm": 0.0, "standoff_bias_mm": 0.0,
"passes": 0}
dt, _v, f = self.intervals()
seg = np.hypot(np.diff(self.x_mm), np.diff(self.z_mm)) / 1000.0
flow_lps = float(self.spray.get("flow_l_min", 0.0)) / 60.0
return {"duration_s": round(float(self.t[-1] - self.t[0]), 2),
"spray_s": round(float(np.sum(f * dt)), 2),
"pump_on_s": round(float(dt[f > 0].sum()), 2),
"paint_l": round(float(np.sum(f * dt) * flow_lps), 3),
"path_m": round(float(seg.sum()), 2),
"spray_path_m": round(float(seg[f > 0].sum()), 2),
"reaction_n": round(float(self.reaction_n), 2), "standoff_sag_mm": round(float(self.sag_mm), 1),
"standoff_bias_mm": round(float(self.standoff_bias_mm), 1),
"passes": int(len({int(p) for p in self.pass_id if p >= 0}))}
def strokes(self):
"""The spraying runs as polylines in the twin's pixel frame (col, row), for the parts that need strokes."""
out, cur = [], []
h = self.shape[0] if self.shape else 0
for i in range(len(self)):
if self.flow[i] > 0:
cur.append([self.x_mm[i] / self.mm_per_px, h - self.z_mm[i] / self.mm_per_px])
elif cur:
if len(cur) >= 2:
out.append(cur)
cur = []
if len(cur) >= 2:
out.append(cur)
return out
def _polyline_mm(stroke, mm_per_px, height_px):
pts = np.asarray(stroke, float)
if pts.ndim != 2 or pts.shape[0] < 2:
raise ValueError("a pass needs at least two points")
return np.column_stack([pts[:, 0] * mm_per_px, (height_px - pts[:, 1]) * mm_per_px])
def _at(pts, cum, s):
"""Position (N, 2) at arc length ``s``, extrapolating along the end segments outside [0, L]."""
s = np.asarray(s, float)
L = float(cum[-1])
out = np.empty((len(s), 2), float)
inside = (s >= 0) & (s <= L)
if inside.any():
out[inside, 0] = np.interp(s[inside], cum, pts[:, 0])
out[inside, 1] = np.interp(s[inside], cum, pts[:, 1])
if (s < 0).any():
d = pts[1] - pts[0]
n = float(np.hypot(*d))
d = d / n if n > 1e-9 else np.array([1.0, 0.0])
out[s < 0] = pts[0] + s[s < 0][:, None] * d # behind the start, along the first segment
if (s > L).any():
d = pts[-1] - pts[-2]
n = float(np.hypot(*d))
d = d / n if n > 1e-9 else np.array([1.0, 0.0])
out[s > L] = pts[-1] + (s[s > L] - L)[:, None] * d # past the end, along the last segment
return out
def from_strokes(strokes_px, mm_per_px, shape, spray, standoff_mm=None, *, speed_ms=None, line_up_s=4.0,
transit_speed_ms=2.0, sortie=1, decel_mm=0.0, end_speed_mm_s=50.0, compensated=True,
lateral_sigma_mm=0.0, seed=0, ds_mm=DS_MM, pressure_psi=2000.0, hold_stiffness_n_per_mm=0.5,
kick_mm=None, kick_tau_s=0.35, compensate_sag=True):
"""A planned sortie as a timeline: each pass at the planned speed and standoff, the gun led by the tip's latency.
``compensated``: the controller issues open ``open_lead`` before the pass start and close ``close_lead`` before its
end (what ``application.trigger_offsets`` exists to tell it), so the fan is at full flow from the first millimetre
of the pass. False models a controller that ignores the lead: the ramps land inside the pass instead.
``decel_mm``: >0 models a controller that slows to ``end_speed_mm_s`` over the last ``decel_mm`` of the pass while
still spraying. That is the pass-end heavy band the audit exists to catch; a controller that closes the gun before
slowing does not produce it (the default).
``hold_stiffness_n_per_mm`` / ``kick_mm``: the spray pushes back (``application.spray_reaction``), so the nozzle sits
further from the wall than the setpoint by force / stiffness, and the trigger adds a decaying kick on top. The
reaction is proportional to flow, so it follows the valve ramps. A stiffer hold (bigger number) sags less; ``kick_mm``
defaults to half the sag, and ``kick_tau_s`` is how long the machine takes to settle.
``compensate_sag``: the plan tells the controller to bias its standoff setpoint in by the sag (the same number
``drone_plan.build`` puts in ``standoff_setpoint_mm`` and the waypoints), so the nozzle holds the design standoff.
``False`` models a controller that ignores the bias: the steady standoff is then the design one *plus* the sag.
"""
h = int(shape[0])
v = float(speed_ms) * 1000.0 if speed_ms is not None else float(spray.get("traverse_speed_mm_s", 400.0))
if v <= 0:
raise ValueError("the pass speed must be positive")
standoff = float(standoff_mm if standoff_mm is not None else spray.get("standoff_mm", 300.0))
leads = application.trigger_offsets(v)
open_lead, close_lead = float(leads["open_lead_mm"]), float(leads["close_lead_mm"])
reaction = application.spray_reaction(float(spray.get("flow_l_min", 0.0)) / 60.0,
float(spray.get("pressure_psi") or pressure_psi))
sag = reaction["reaction_n"] / max(float(hold_stiffness_n_per_mm), 1e-3)
kick = (0.5 * sag) if kick_mm is None else float(kick_mm)
bias = sag if compensate_sag else 0.0 # the setpoint the controller is asked to hold
base = standoff - bias
rng = np.random.default_rng(seed)
step = max(2.0, min(float(ds_mm), 2.0 * float(mm_per_px)))
T, X, Z, S, V, F, P, PH = [], [], [], [], [], [], [], []
t_cursor, prev_end, pid = 0.0, None, 0
for stroke in strokes_px:
pts = _polyline_mm(stroke, mm_per_px, h)
seg = np.hypot(*np.diff(pts, axis=0).T)
cum = np.concatenate([[0.0], np.cumsum(seg)])
L = float(cum[-1])
if L < 1.0:
continue
pid += 1
s0 = -open_lead if compensated else 0.0
s1 = L + close_lead if compensated else L
n = max(8, int(math.ceil((s1 - s0) / step)) + 1)
s = np.linspace(s0, s1, n)
xy = _at(pts, cum, s)
x, z = xy[:, 0], xy[:, 1]
if lateral_sigma_mm > 0: # what-if: this pass flown off its line
err = rng.normal(0.0, float(lateral_sigma_mm))
if abs(x[-1] - x[0]) >= abs(z[-1] - z[0]):
z = z + err
else:
x = x + err
speed = np.full(n, v, float)
if decel_mm > 0: # the controller slows while the gun is still open
tail = s > L - float(decel_mm)
k = np.clip((L - s) / float(decel_mm), 0.0, 1.0)
speed = np.where(tail, float(end_speed_mm_s) + (v - float(end_speed_mm_s)) * k, v)
speed = np.maximum(speed, 1.0)
if compensated:
flow = np.clip((s + open_lead) / max(open_lead, 1e-6), 0, 1) * np.clip((L + close_lead - s) / max(close_lead, 1e-6), 0, 1)
else:
flow = np.clip(s / max(open_lead, 1e-6), 0, 1) * np.clip((L - s) / max(close_lead, 1e-6), 0, 1)
flow = np.clip(flow, 0.0, 1.0)
if prev_end is not None:
t_cursor += float(np.hypot(*(xy[0] - prev_end))) / (float(transit_speed_ms) * 1000.0)
T.append(t_cursor); X.append(xy[0, 0]); Z.append(xy[0, 1]); S.append(base)
V.append(0.0); F.append(0.0); P.append(-1); PH.append("lineup")
t_cursor += float(line_up_s)
ds = np.diff(s)
dt = ds / (0.5 * (speed[:-1] + speed[1:])) if n > 1 else np.zeros(0)
tp = t_cursor + np.concatenate([[0.0], np.cumsum(dt)])
off = sag * flow # the jet pushes the nozzle off its standoff...
if kick > 0: # ...and the trigger adds a decaying kick on top
off = off + kick * flow * np.exp(-np.maximum(tp - tp[0], 0.0) / max(float(kick_tau_s), 1e-3))
T.append(tp); X.append(x); Z.append(z); S.append(np.full(n, base) + off)
V.append(speed); F.append(flow); P.append(np.full(n, pid))
PH.append(np.where(flow > 0, "spray", "transit"))
t_cursor = float(tp[-1])
prev_end = xy[-1]
if not T: # a plan with nothing to fly is a valid (empty) sortie, not an error
empty = np.zeros(0, np.float64)
return Timeline(t=empty, x_mm=empty.astype(np.float32), z_mm=empty.astype(np.float32),
standoff_mm=empty.astype(np.float32), speed_mm_s=empty.astype(np.float32),
flow=empty.astype(np.float32), pass_id=np.zeros(0, np.int32), phase=np.zeros(0, "<U8"),
sortie=int(sortie), source="planned", spray=dict(spray), mm_per_px=float(mm_per_px),
shape=(h, int(shape[1])), reaction_n=reaction["reaction_n"], sag_mm=sag, standoff_bias_mm=bias)
cat = lambda k, dt_: np.concatenate([np.atleast_1d(x) for x in k]).astype(dt_)
return Timeline(t=cat(T, np.float64), x_mm=cat(X, np.float32), z_mm=cat(Z, np.float32),
standoff_mm=cat(S, np.float32), speed_mm_s=cat(V, np.float32), flow=cat(F, np.float32),
pass_id=cat(P, np.int32), phase=cat(PH, "<U8"), sortie=int(sortie), source="planned",
spray=dict(spray), mm_per_px=float(mm_per_px), shape=(h, int(shape[1])),
reaction_n=reaction["reaction_n"], sag_mm=sag, standoff_bias_mm=bias)
def from_log(log, a_en, b_en, mm_per_px, height_px, ground_alt_m=0.0, target_standoff_m=0.3, boom_m=0.5,
spray=None, *, sortie=1, width_px=None, speed_window=5, max_gap_s=1.5, pressure_psi=2000.0):
"""A flight log as a timeline: the samples the aircraft reported, with the pump state as the valve flow.
Speed is the distance between samples over their time step, smoothed over ``speed_window`` samples; a pump-on run
ends when the pump goes off or the samples skip more than ``max_gap_s``. That is the point of this constructor: a log
already contains the wobble, the standoff drift and the timing, so nothing here has to be assumed about them.
The spray's push on the nozzle is still a design quantity (flow and pressure, from ``spray``), so it is computed
here; the standoff *cost* is not modelled but measured - the median standoff while spraying against the design one.
"""
import drone_log as dlog
w = dlog.to_wall_frame(log, a_en, b_en)
t = np.asarray(w["t"], float)
if len(t) < 2:
raise ValueError("the log has fewer than 2 samples")
x = np.asarray(w["x"], float) * 1000.0
z = (np.asarray(w["z"], float) - float(ground_alt_m)) * 1000.0
sd = np.maximum(np.asarray(w["y"], float) * 1000.0 - float(boom_m) * 1000.0, 10.0)
pump = np.asarray(w["pump"], bool)
flow_lps = float((spray or {}).get("flow_l_min", 0.0)) / 60.0
reaction_n = float(application.spray_reaction(flow_lps, float((spray or {}).get("pressure_psi") or pressure_psi))["reaction_n"]) if flow_lps > 0 else 0.0
sag_mm = float(np.median(sd[pump]) - float(target_standoff_m) * 1000.0) if pump.any() and flow_lps > 0 else 0.0
dt = np.diff(t, prepend=t[0])
dist = np.hypot(np.diff(x, prepend=x[0]), np.diff(z, prepend=z[0]))
v = dist / np.maximum(dt, 1e-6)
v[0] = v[1] if len(v) > 1 else 0.0
if speed_window > 1 and len(v) >= speed_window:
n = int(speed_window)
half = n // 2
v = np.convolve(np.pad(v, (half, n - 1 - half), mode="edge"), np.ones(n) / n, mode="valid")
pid = np.full(len(t), -1, int)
current, prev, last_on = 0, False, None
for i in range(len(t)):
on = bool(pump[i])
if on and (not prev or (last_on is not None and t[i] - last_on > float(max_gap_s))):
current += 1
if on:
pid[i] = current
last_on = t[i]
prev = on
return Timeline(t=t, x_mm=x.astype(np.float32), z_mm=z.astype(np.float32), standoff_mm=sd.astype(np.float32),
speed_mm_s=np.maximum(v, 0.0).astype(np.float32), flow=pump.astype(np.float32),
pass_id=pid.astype(np.int32), phase=np.where(pump, "spray", "transit").astype("<U8"),
sortie=int(sortie), source="log", spray=dict(spray or {}), mm_per_px=float(mm_per_px),
shape=(int(height_px), int(width_px) if width_px else 0), ramps=False,
reaction_n=reaction_n, sag_mm=sag_mm)
|