CHEVROLET BOLT EUV 2022: Add a simple neural feed forward (#31266)

* add simple neural feed forward

* update refs

* do not sample during inference in op

* update refs
old-commit-hash: 619625625c

selfdrive/car/gm/interface.py

@@ -1,13 +1,15 @@
 #!/usr/bin/env python3
+import os
 from cereal import car
 from math import fabs, exp
 from panda import Panda
 
+from openpilot.common.basedir import BASEDIR
 from openpilot.common.conversions import Conversions as CV
 from openpilot.selfdrive.car import create_button_events, get_safety_config
 from openpilot.selfdrive.car.gm.radar_interface import RADAR_HEADER_MSG
 from openpilot.selfdrive.car.gm.values import CAR, CruiseButtons, CarControllerParams, EV_CAR, CAMERA_ACC_CAR, CanBus
-from openpilot.selfdrive.car.interfaces import CarInterfaceBase, TorqueFromLateralAccelCallbackType, FRICTION_THRESHOLD, LatControlInputs
+from openpilot.selfdrive.car.interfaces import CarInterfaceBase, TorqueFromLateralAccelCallbackType, FRICTION_THRESHOLD, LatControlInputs, NanoFFModel
 from openpilot.selfdrive.controls.lib.drive_helpers import get_friction
 
 ButtonType = car.CarState.ButtonEvent.Type
@@ -25,6 +27,8 @@ NON_LINEAR_TORQUE_PARAMS = {
   CAR.SILVERADO: [3.29974374, 1.0, 0.25571356, 0.0465122]
 }
 
+NEURAL_PARAMS_PATH = os.path.join(BASEDIR, 'selfdrive/car/torque_data/neural_ff_weights.json')
+
 
 class CarInterface(CarInterfaceBase):
   @staticmethod
@@ -61,8 +65,19 @@ class CarInterface(CarInterfaceBase):
     steer_torque = (sig(latcontrol_inputs.lateral_acceleration * a) * b) + (latcontrol_inputs.lateral_acceleration * c)
     return float(steer_torque) + friction
 
+  def torque_from_lateral_accel_neural(self, latcontrol_inputs: LatControlInputs, torque_params: car.CarParams.LateralTorqueTuning, lateral_accel_error: float,
+                                       lateral_accel_deadzone: float, friction_compensation: bool, gravity_adjusted: bool) -> float:
+    friction = get_friction(lateral_accel_error, lateral_accel_deadzone, FRICTION_THRESHOLD, torque_params, friction_compensation)
+    inputs = list(latcontrol_inputs)
+    if gravity_adjusted:
+      inputs[0] += inputs[1]
+    return float(self.neural_ff_model.predict(inputs)) + friction
+
   def torque_from_lateral_accel(self) -> TorqueFromLateralAccelCallbackType:
-    if self.CP.carFingerprint in NON_LINEAR_TORQUE_PARAMS:
+    if self.CP.carFingerprint == CAR.BOLT_EUV:
+      self.neural_ff_model = NanoFFModel(NEURAL_PARAMS_PATH, self.CP.carFingerprint)
+      return self.torque_from_lateral_accel_neural
+    elif self.CP.carFingerprint in NON_LINEAR_TORQUE_PARAMS:
       return self.torque_from_lateral_accel_siglin
     else:
       return self.torque_from_lateral_accel_linear

selfdrive/car/interfaces.py

@@ -1,3 +1,4 @@
+import json
 import os
 import time
 import numpy as np
@@ -473,3 +474,35 @@ def get_interface_attr(attr: str, combine_brands: bool = False, ignore_none: boo
       pass
 
   return result
+
+
+class NanoFFModel:
+  def __init__(self, weights_loc: str, platform: str):
+    self.weights_loc = weights_loc
+    self.platform = platform
+    self.load_weights(platform)
+
+  def load_weights(self, platform: str):
+    with open(self.weights_loc, 'r') as fob:
+      self.weights = {k: np.array(v) for k, v in json.load(fob)[platform].items()}
+
+  def relu(self, x: np.ndarray):
+    return np.maximum(0.0, x)
+
+  def forward(self, x: np.ndarray):
+    assert x.ndim == 1
+    x = (x - self.weights['input_norm_mat'][:, 0]) / (self.weights['input_norm_mat'][:, 1] - self.weights['input_norm_mat'][:, 0])
+    x = self.relu(np.dot(x, self.weights['w_1']) + self.weights['b_1'])
+    x = self.relu(np.dot(x, self.weights['w_2']) + self.weights['b_2'])
+    x = self.relu(np.dot(x, self.weights['w_3']) + self.weights['b_3'])
+    x = np.dot(x, self.weights['w_4']) + self.weights['b_4']
+    return x
+
+  def predict(self, x: List[float], do_sample: bool = False):
+    x = self.forward(np.array(x))
+    if do_sample:
+      pred = np.random.laplace(x[0], np.exp(x[1]) / self.weights['temperature'])
+    else:
+      pred = x[0]
+    pred = pred * (self.weights['output_norm_mat'][1] - self.weights['output_norm_mat'][0]) + self.weights['output_norm_mat'][0]
+    return pred

selfdrive/test/process_replay/ref_commit

@@ -1 +1 @@
-eaab6bd55c5eab33fc9a0d8de8289b912e923887
+d9a3a0d4e806b49ec537233d30926bec70308485