Reapply "Online lateral lag learning" (#34975)

* Online lateral lag learning (#34974)

This reverts commit b4cc9e6.

* pad to the best size for fft

* Fix static analysis

* Add typing

* Fix typing

* MAX_LAG

* Calculate cross correlation regardless if the points are valid

* Back to lagd

* Add lagd to process_config

* Lagd in test onroad

* Move lag estimator for lagd

* Remove duplicate entry from test_onroad

* Update process replay

* pre-fill the data

* Update cpu usage

* 25sec window

* Change the meaning of lateralDelayEstimate

* No newline

* Fix typing

* Prefill

* Update ref commit

* Add a unit test

* Fix static issues

* Time limit

* Or timeout

* Use mocker

* Update estimate every time

* empty test

* DT const

* enable RIVIAN again

* Update ref commit

* Update that again

* Improve the tests

* Fix static

* Add masking test

* Increase timeout

* Add liveDelay to selfdrived

* Add liveDelay to selfdrived in process_replay

* Fix block_avg restore after num_blocks

* regen most

* Update bolt

* Update ref commit

* Change the key name

* Add assert

* True weighted average

Author: fredyshox

cereal/services.py

@@ -36,6 +36,7 @@ def __init__(self, should_log: bool, frequency: float, decimation: Optional[int]
   "errorLogMessage": (True, 0., 1),
   "liveCalibration": (True, 4., 4),
   "liveTorqueParameters": (True, 4., 1),
+  "liveDelay": (True, 4., 1),
   "androidLog": (True, 0.),
   "carState": (True, 100., 10),
   "carControl": (True, 100., 10),

common/params_keys.h

@@ -71,6 +71,7 @@ inline static std::unordered_map<std::string, uint32_t> keys = {
     {"LastPowerDropDetected", CLEAR_ON_MANAGER_START},
     {"LastUpdateException", CLEAR_ON_MANAGER_START},
     {"LastUpdateTime", PERSISTENT},
+    {"LiveDelay", PERSISTENT},
     {"LiveParameters", PERSISTENT},
     {"LiveParametersV2", PERSISTENT},
     {"LiveTorqueParameters", PERSISTENT | DONT_LOG},

selfdrive/locationd/helpers.py

@@ -1,10 +1,48 @@
 import numpy as np
 from typing import Any
+from functools import cache
 
 from cereal import log
 from openpilot.common.transformations.orientation import rot_from_euler, euler_from_rot
 
 
+@cache
+def fft_next_good_size(n: int) -> int:
+    """
+    smallest composite of 2, 3, 5, 7, 11 that is >= n
+    inspired by pocketfft
+    """
+    if n <= 6:
+      return n
+    best, f2 = 2 * n, 1
+    while f2 < best:
+        f23 = f2
+        while f23 < best:
+            f235 = f23
+            while f235 < best:
+                f2357 = f235
+                while f2357 < best:
+                    f235711 = f2357
+                    while f235711 < best:
+                        best = f235711 if f235711 >= n else best
+                        f235711 *= 11
+                    f2357 *= 7
+                f235 *= 5
+            f23 *= 3
+        f2 *= 2
+    return best
+
+
+def parabolic_peak_interp(R, max_index):
+  if max_index == 0 or max_index == len(R) - 1:
+    return max_index
+
+  y_m1, y_0, y_p1 = R[max_index - 1], R[max_index], R[max_index + 1]
+  offset = 0.5 * (y_p1 - y_m1) / (2 * y_0 - y_p1 - y_m1)
+
+  return max_index + offset
+
+
 def rotate_cov(rot_matrix, cov_in):
   return rot_matrix @ cov_in @ rot_matrix.T
 

selfdrive/locationd/lagd.py

@@ -0,0 +1,333 @@
+#!/usr/bin/env python3
+import os
+import numpy as np
+import capnp
+from collections import deque
+from functools import partial
+
+import cereal.messaging as messaging
+from cereal import car, log
+from cereal.services import SERVICE_LIST
+from openpilot.common.params import Params
+from openpilot.common.realtime import config_realtime_process
+from openpilot.common.swaglog import cloudlog
+from openpilot.selfdrive.locationd.helpers import PoseCalibrator, Pose, fft_next_good_size, parabolic_peak_interp
+
+BLOCK_SIZE = 100
+BLOCK_NUM = 50
+BLOCK_NUM_NEEDED = 5
+MOVING_WINDOW_SEC = 300.0
+MIN_OKAY_WINDOW_SEC = 25.0
+MIN_RECOVERY_BUFFER_SEC = 2.0
+MIN_VEGO = 15.0
+MIN_ABS_YAW_RATE = np.radians(1.0)
+MIN_NCC = 0.95
+MAX_LAG = 1.0
+
+
+def masked_normalized_cross_correlation(expected_sig: np.ndarray, actual_sig: np.ndarray, mask: np.ndarray, n: int):
+  """
+  References:
+    D. Padfield. "Masked FFT registration". In Proc. Computer Vision and
+    Pattern Recognition, pp. 2918-2925 (2010).
+    :DOI:`10.1109/CVPR.2010.5540032`
+  """
+
+  eps = np.finfo(np.float64).eps
+  expected_sig = np.asarray(expected_sig, dtype=np.float64)
+  actual_sig = np.asarray(actual_sig, dtype=np.float64)
+
+  expected_sig[~mask] = 0.0
+  actual_sig[~mask] = 0.0
+
+  rotated_expected_sig = expected_sig[::-1]
+  rotated_mask = mask[::-1]
+
+  fft = partial(np.fft.fft, n=n)
+
+  actual_sig_fft = fft(actual_sig)
+  rotated_expected_sig_fft = fft(rotated_expected_sig)
+  actual_mask_fft = fft(mask.astype(np.float64))
+  rotated_mask_fft = fft(rotated_mask.astype(np.float64))
+
+  number_overlap_masked_samples = np.fft.ifft(rotated_mask_fft * actual_mask_fft).real
+  number_overlap_masked_samples[:] = np.round(number_overlap_masked_samples)
+  number_overlap_masked_samples[:] = np.fmax(number_overlap_masked_samples, eps)
+  masked_correlated_actual_fft = np.fft.ifft(rotated_mask_fft * actual_sig_fft).real
+  masked_correlated_expected_fft = np.fft.ifft(actual_mask_fft * rotated_expected_sig_fft).real
+
+  numerator = np.fft.ifft(rotated_expected_sig_fft * actual_sig_fft).real
+  numerator -= masked_correlated_actual_fft * masked_correlated_expected_fft / number_overlap_masked_samples
+
+  actual_squared_fft = fft(actual_sig ** 2)
+  actual_sig_denom = np.fft.ifft(rotated_mask_fft * actual_squared_fft).real
+  actual_sig_denom -= masked_correlated_actual_fft ** 2 / number_overlap_masked_samples
+  actual_sig_denom[:] = np.fmax(actual_sig_denom, 0.0)
+
+  rotated_expected_squared_fft = fft(rotated_expected_sig ** 2)
+  expected_sig_denom = np.fft.ifft(actual_mask_fft * rotated_expected_squared_fft).real
+  expected_sig_denom -= masked_correlated_expected_fft ** 2 / number_overlap_masked_samples
+  expected_sig_denom[:] = np.fmax(expected_sig_denom, 0.0)
+
+  denom = np.sqrt(actual_sig_denom * expected_sig_denom)
+
+  # zero-out samples with very small denominators
+  tol = 1e3 * eps * np.max(np.abs(denom), keepdims=True)
+  nonzero_indices = denom > tol
+
+  ncc = np.zeros_like(denom, dtype=np.float64)
+  ncc[nonzero_indices] = numerator[nonzero_indices] / denom[nonzero_indices]
+  np.clip(ncc, -1, 1, out=ncc)
+
+  return ncc
+
+
+class Points:
+  def __init__(self, num_points: int):
+    self.times = deque[float]([0.0] * num_points, maxlen=num_points)
+    self.okay = deque[bool]([False] * num_points, maxlen=num_points)
+    self.desired = deque[float]([0.0] * num_points, maxlen=num_points)
+    self.actual = deque[float]([0.0] * num_points, maxlen=num_points)
+
+  @property
+  def num_points(self):
+    return len(self.desired)
+
+  @property
+  def num_okay(self):
+    return np.count_nonzero(self.okay)
+
+  def update(self, t: float, desired: float, actual: float, okay: bool):
+    self.times.append(t)
+    self.okay.append(okay)
+    self.desired.append(desired)
+    self.actual.append(actual)
+
+  def get(self) -> tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
+    return np.array(self.times), np.array(self.desired), np.array(self.actual), np.array(self.okay)
+
+
+class BlockAverage:
+  def __init__(self, num_blocks: int, block_size: int, valid_blocks: int, initial_value: float):
+    self.num_blocks = num_blocks
+    self.block_size = block_size
+    self.block_idx = valid_blocks % num_blocks
+    self.idx = 0
+
+    self.values = np.tile(initial_value, (num_blocks, 1))
+    self.valid_blocks = valid_blocks
+
+  def update(self, value: float):
+    self.values[self.block_idx] = (self.idx * self.values[self.block_idx] + value) / (self.idx + 1)
+    self.idx = (self.idx + 1) % self.block_size
+    if self.idx == 0:
+      self.block_idx = (self.block_idx + 1) % self.num_blocks
+      self.valid_blocks = min(self.valid_blocks + 1, self.num_blocks)
+
+  def get(self) -> tuple[float, float]:
+    valid_block_idx = [i for i in range(self.valid_blocks) if i != self.block_idx]
+    valid_and_current_idx = valid_block_idx + ([self.block_idx] if self.idx > 0 else [])
+
+    valid_mean = float(np.mean(self.values[valid_block_idx], axis=0).item()) if len(valid_block_idx) > 0 else float('nan')
+    current_mean = float(np.mean(self.values[valid_and_current_idx], axis=0).item()) if len(valid_and_current_idx) > 0 else float('nan')
+    return valid_mean, current_mean
+
+
+class LateralLagEstimator:
+  inputs = {"carControl", "carState", "controlsState", "liveCalibration", "livePose"}
+
+  def __init__(self, CP: car.CarParams, dt: float,
+               block_count: int = BLOCK_NUM, min_valid_block_count: int = BLOCK_NUM_NEEDED, block_size: int = BLOCK_SIZE,
+               window_sec: float = MOVING_WINDOW_SEC, okay_window_sec: float = MIN_OKAY_WINDOW_SEC, min_recovery_buffer_sec: float = MIN_RECOVERY_BUFFER_SEC,
+               min_vego: float = MIN_VEGO, min_yr: float = MIN_ABS_YAW_RATE, min_ncc: float = MIN_NCC):
+    self.dt = dt
+    self.window_sec = window_sec
+    self.okay_window_sec = okay_window_sec
+    self.min_recovery_buffer_sec = min_recovery_buffer_sec
+    self.initial_lag = CP.steerActuatorDelay + 0.2
+    self.block_size = block_size
+    self.block_count = block_count
+    self.min_valid_block_count = min_valid_block_count
+    self.min_vego = min_vego
+    self.min_yr = min_yr
+    self.min_ncc = min_ncc
+
+    self.t = 0.0
+    self.lat_active = False
+    self.steering_pressed = False
+    self.steering_saturated = False
+    self.desired_curvature = 0.0
+    self.v_ego = 0.0
+    self.yaw_rate = 0.0
+
+    self.last_lat_inactive_t = 0.0
+    self.last_steering_pressed_t = 0.0
+    self.last_steering_saturated_t = 0.0
+    self.last_estimate_t = 0.0
+
+    self.calibrator = PoseCalibrator()
+
+    self.reset(self.initial_lag, 0)
+
+  def reset(self, initial_lag: float, valid_blocks: int):
+    window_len = int(self.window_sec / self.dt)
+    self.points = Points(window_len)
+    self.block_avg = BlockAverage(self.block_count, self.block_size, valid_blocks, initial_lag)
+
+  def get_msg(self, valid: bool, debug: bool = False) -> capnp._DynamicStructBuilder:
+    msg = messaging.new_message('liveDelay')
+
+    msg.valid = valid
+
+    liveDelay = msg.liveDelay
+
+    valid_mean_lag, current_mean_lag = self.block_avg.get()
+    if self.block_avg.valid_blocks >= self.min_valid_block_count and not np.isnan(valid_mean_lag):
+      liveDelay.status = log.LiveDelayData.Status.estimated
+      liveDelay.lateralDelay = valid_mean_lag
+    else:
+      liveDelay.status = log.LiveDelayData.Status.unestimated
+      liveDelay.lateralDelay = self.initial_lag
+    if not np.isnan(current_mean_lag):
+      liveDelay.lateralDelayEstimate = current_mean_lag
+    else:
+      liveDelay.lateralDelayEstimate = self.initial_lag
+    liveDelay.validBlocks = self.block_avg.valid_blocks
+    if debug:
+      liveDelay.points = self.block_avg.values.flatten().tolist()
+
+    return msg
+
+  def handle_log(self, t: float, which: str, msg: capnp._DynamicStructReader):
+    if which == "carControl":
+      self.lat_active = msg.latActive
+    elif which == "carState":
+      self.steering_pressed = msg.steeringPressed
+      self.v_ego = msg.vEgo
+    elif which == "controlsState":
+      self.steering_saturated = getattr(msg.lateralControlState, msg.lateralControlState.which()).saturated
+      self.desired_curvature = msg.desiredCurvature
+    elif which == "liveCalibration":
+      self.calibrator.feed_live_calib(msg)
+    elif which == "livePose":
+      device_pose = Pose.from_live_pose(msg)
+      calibrated_pose = self.calibrator.build_calibrated_pose(device_pose)
+      self.yaw_rate = calibrated_pose.angular_velocity.z
+    self.t = t
+
+  def points_enough(self):
+    return self.points.num_points >= int(self.okay_window_sec / self.dt)
+
+  def points_valid(self):
+    return self.points.num_okay >= int(self.okay_window_sec / self.dt)
+
+  def update_points(self):
+    if not self.lat_active:
+      self.last_lat_inactive_t = self.t
+    if self.steering_pressed:
+      self.last_steering_pressed_t = self.t
+    if self.steering_saturated:
+      self.last_steering_saturated_t = self.t
+
+    la_desired = self.desired_curvature * self.v_ego * self.v_ego
+    la_actual_pose = self.yaw_rate * self.v_ego
+
+    fast = self.v_ego > self.min_vego
+    turning = np.abs(self.yaw_rate) >= self.min_yr
+    has_recovered = all( # wait for recovery after !lat_active, steering_pressed, steering_saturated
+      self.t - last_t >= self.min_recovery_buffer_sec
+      for last_t in [self.last_lat_inactive_t, self.last_steering_pressed_t, self.last_steering_saturated_t]
+    )
+    okay = self.lat_active and not self.steering_pressed and not self.steering_saturated and fast and turning and has_recovered
+
+    self.points.update(self.t, la_desired, la_actual_pose, okay)
+
+  def update_estimate(self):
+    if not self.points_enough():
+      return
+
+    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()
+    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:]))
+
+    delay, corr = self.actuator_delay(desired, actual, okay, self.dt, MAX_LAG)
+    if corr < self.min_ncc or not is_valid:
+      return
+
+    self.block_avg.update(delay)
+    self.last_estimate_t = self.t
+
+  def actuator_delay(self, expected_sig: np.ndarray, actual_sig: np.ndarray, mask: np.ndarray, dt: float, max_lag: float) -> tuple[float, float]:
+    assert len(expected_sig) == len(actual_sig)
+    max_lag_samples = int(max_lag / dt)
+    padded_size = fft_next_good_size(len(expected_sig) + max_lag_samples)
+
+    ncc = masked_normalized_cross_correlation(expected_sig, actual_sig, mask, padded_size)
+
+    # only consider lags from 0 to max_lag
+    roi_ncc = ncc[len(expected_sig) - 1: len(expected_sig) - 1 + max_lag_samples]
+
+    max_corr_index = np.argmax(roi_ncc)
+    corr = roi_ncc[max_corr_index]
+    lag = parabolic_peak_interp(roi_ncc, max_corr_index) * dt
+
+    return lag, corr
+
+
+def retrieve_initial_lag(params_reader: Params, CP: car.CarParams):
+  last_lag_data = params_reader.get("LiveDelay")
+  last_carparams_data = params_reader.get("CarParamsPrevRoute")
+
+  if last_lag_data is not None:
+    try:
+      with log.Event.from_bytes(last_lag_data) as last_lag_msg, car.CarParams.from_bytes(last_carparams_data) as last_CP:
+        ld = last_lag_msg.liveDelay
+        if last_CP.carFingerprint != CP.carFingerprint:
+          raise Exception("Car model mismatch")
+
+        lag, valid_blocks = ld.lateralDelayEstimate, ld.validBlocks
+        assert valid_blocks <= BLOCK_NUM, "Invalid number of valid blocks"
+        return lag, valid_blocks
+    except Exception as e:
+      cloudlog.error(f"Failed to retrieve initial lag: {e}")
+
+  return None
+
+
+def main():
+  config_realtime_process([0, 1, 2, 3], 5)
+
+  DEBUG = bool(int(os.getenv("DEBUG", "0")))
+
+  pm = messaging.PubMaster(['liveDelay'])
+  sm = messaging.SubMaster(['livePose', 'liveCalibration', 'carState', 'controlsState', 'carControl'], poll='livePose')
+
+  params_reader = Params()
+  CP = messaging.log_from_bytes(params_reader.get("CarParams", block=True), car.CarParams)
+
+  lag_learner = LateralLagEstimator(CP, 1. / SERVICE_LIST['livePose'].frequency)
+  if (initial_lag_params := retrieve_initial_lag(params_reader, CP)) is not None:
+    lag, valid_blocks = initial_lag_params
+    lag_learner.reset(lag, valid_blocks)
+
+  while True:
+    sm.update()
+    if sm.all_checks():
+      for which in sorted(sm.updated.keys(), key=lambda x: sm.logMonoTime[x]):
+        if sm.updated[which]:
+          t = sm.logMonoTime[which] * 1e-9
+          lag_learner.handle_log(t, which, sm[which])
+      lag_learner.update_points()
+
+    # 4Hz driven by livePose
+    if sm.frame % 5 == 0:
+      lag_learner.update_estimate()
+      lag_msg = lag_learner.get_msg(sm.all_checks(), DEBUG)
+      lag_msg_dat = lag_msg.to_bytes()
+      pm.send('liveDelay', lag_msg_dat)
+
+      if sm.frame % 1200 == 0: # cache every 60 seconds
+        params_reader.put_nonblocking("LiveDelay", lag_msg_dat)