From 5e2a5b53553dab1fbacf964aac695c9c4c7e2df0 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Kacper=20R=C4=85czy?= <gfw.kra@gmail.com>
Date: Fri, 6 Mar 2026 19:00:15 -0800
Subject: [PATCH] lagd: smooth lat accel + min lat accel range (#37424)

* Smooth

* Min lat accel range

* Make the moving average masked

* Bring back the range

* Update test

* Smooth desired signal too

* Diff

* Gaussian

* Fix fmt

* Remove newline
---
 selfdrive/locationd/lagd.py           | 20 +++++++++++++++++++-
 selfdrive/locationd/test/test_lagd.py |  4 ++--
 2 files changed, 21 insertions(+), 3 deletions(-)

diff --git a/selfdrive/locationd/lagd.py b/selfdrive/locationd/lagd.py
index 361bb79cc..8bbee6604 100755
--- a/selfdrive/locationd/lagd.py
+++ b/selfdrive/locationd/lagd.py
@@ -28,11 +28,26 @@ MIN_LAG = 0.15
 MAX_LAG_STD = 0.1
 MAX_LAT_ACCEL = 2.0
 MAX_LAT_ACCEL_DIFF = 0.6
+MIN_LAT_ACCEL_RANGE = 0.5
 MIN_CONFIDENCE = 0.7
 CORR_BORDER_OFFSET = 5
 LAG_CANDIDATE_CORR_THRESHOLD = 0.9
+SMOOTH_K = 5
+SMOOTH_SIGMA = 1.0
 
 
+def masked_symmetric_moving_average(x: np.ndarray, mask: np.ndarray, k: int, sigma: float) -> np.ndarray:
+  assert k >= 1 and k % 2 == 1, "k must be positive and odd"
+  pad = k // 2
+  i = np.arange(k) - pad
+  w = np.exp(-0.5 * (i / sigma) ** 2)
+  w /= w.sum()
+  xp = np.pad(x * mask, pad, mode="edge")
+  mp = np.pad(mask, pad, mode="edge")
+  num = np.convolve(xp, w, mode="valid")
+  den = np.convolve(mp, w, mode="valid")
+  return np.divide(num, den, out=np.full_like(num, np.nan, dtype=np.float64), where=den != 0)
+
 def masked_normalized_cross_correlation(expected_sig: np.ndarray, actual_sig: np.ndarray, mask: np.ndarray, n: int):
   """
   References:
@@ -294,11 +309,14 @@ class LateralLagEstimator:
 
     times, desired, actual, okay = self.points.get()
     # check if there are any new valid data points since the last update
-    is_valid = self.points_valid()
+    is_valid = self.points_valid() and (actual.max() - actual.min() >= MIN_LAT_ACCEL_RANGE)
     if self.last_estimate_t != 0 and times[0] <= self.last_estimate_t:
       new_values_start_idx = next(-i for i, t in enumerate(reversed(times)) if t <= self.last_estimate_t)
       is_valid = is_valid and not (new_values_start_idx == 0 or not np.any(okay[new_values_start_idx:]))
 
+    desired = masked_symmetric_moving_average(desired, okay, SMOOTH_K, SMOOTH_SIGMA)
+    actual = masked_symmetric_moving_average(actual, okay, SMOOTH_K, SMOOTH_SIGMA)
+
     delay, corr, confidence = self.actuator_delay(desired, actual, okay, self.dt, MIN_LAG, MAX_LAG)
     if corr < self.min_ncc or confidence < self.min_confidence or not is_valid:
       return
diff --git a/selfdrive/locationd/test/test_lagd.py b/selfdrive/locationd/test/test_lagd.py
index 4728413d9..6249e6b04 100644
--- a/selfdrive/locationd/test/test_lagd.py
+++ b/selfdrive/locationd/test/test_lagd.py
@@ -19,8 +19,8 @@ DT = 0.05
 def process_messages(estimator, lag_frames, n_frames, vego=20.0, rejection_threshold=0.0):
   for i in range(n_frames):
     t = i * estimator.dt
-    desired_la = np.cos(10 * t) * 0.1
-    actual_la = np.cos(10 * (t - lag_frames * estimator.dt)) * 0.1
+    desired_la = np.cos(10 * t) * 0.3
+    actual_la = np.cos(10 * (t - lag_frames * estimator.dt)) * 0.3
 
     # if sample is masked out, set it to desired value (no lag)
     rejected = random.uniform(0, 1) < rejection_threshold