yaw_est: use error-state covariance prediction

Convergence improvements in high yaw rate conditions

src/modules/ekf2/EKF/EKFGSF_yaw.cpp

@@ -283,6 +283,7 @@ void EKFGSF_yaw::predictEKF(const uint8_t model_index, const Vector3f &delta_ang
 	const float sin_yaw = sinf(_ekf_gsf[model_index].X(2));
 	const float dvx =   del_vel_NED(0) * cos_yaw + del_vel_NED(1) * sin_yaw;
 	const float dvy = - del_vel_NED(0) * sin_yaw + del_vel_NED(1) * cos_yaw;
+	const float daz = Vector3f(_ahrs_ekf_gsf[model_index].R * delta_ang)(2);
 
 	// delta velocity process noise double if we're not in air
 	const float accel_noise = in_air ? _accel_noise : 2.f * _accel_noise;
@@ -292,7 +293,7 @@ void EKFGSF_yaw::predictEKF(const uint8_t model_index, const Vector3f &delta_ang
 	const float d_ang_var = sq(_gyro_noise * delta_ang_dt);
 
 	sym::YawEstPredictCovariance(_ekf_gsf[model_index].X, _ekf_gsf[model_index].P,
-				     Vector2f(dvx, dvy), d_vel_var, d_ang_var, &_ekf_gsf[model_index].P);
+				     Vector2f(dvx, dvy), d_vel_var, daz, d_ang_var, &_ekf_gsf[model_index].P);
 
 	// covariance matrix is symmetrical, so copy upper half to lower half
 	_ekf_gsf[model_index].P(1, 0) = _ekf_gsf[model_index].P(0, 1);

src/modules/ekf2/EKF/python/ekf_derivation/derivation_yaw_estimator.py

@@ -1,7 +1,7 @@
 #!/usr/bin/env python
 # -*- coding: utf-8 -*-
 """
-    Copyright (c) 2022 PX4 Development Team
+    Copyright (c) 2022-2023 PX4 Development Team
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions
     are met:
@@ -37,51 +37,86 @@
 symforce.set_epsilon_to_symbol()
 
 import symforce.symbolic as sf
+from symforce.values import Values
 from derivation_utils import *
 
-class State:
-    vx = 0
-    vy = 1
-    yaw = 2
-    n_states = 3
+State = Values(
+    vel = sf.V2(),
+    R = sf.Rot2() # 2D rotation to handle angle wrap
+)
 
-class VState(sf.Matrix):
-    SHAPE = (State.n_states, 1)
+class VTangent(sf.Matrix):
+    SHAPE = (State.tangent_dim(), 1)
 
-class MState(sf.Matrix):
-    SHAPE = (State.n_states, State.n_states)
+class MTangent(sf.Matrix):
+    SHAPE = (State.tangent_dim(), State.tangent_dim())
+
+def rot2_small_angle(angle: sf.V1):
+    # Approximation for small "delta angles" to avoid trigonometric functions
+    return sf.Rot2(sf.Complex(1, angle[0]))
 
 def yaw_est_predict_covariance(
-        state: VState,
-        P: MState,
+        state: VTangent,
+        P: MTangent,
         d_vel: sf.V2,
         d_vel_var: sf.Scalar,
-        d_ang_var: sf.Scalar
+        d_ang: sf.Scalar,
+        d_ang_var: sf.Scalar,
 ):
-    d_ang = sf.Symbol("d_ang") # does not appear in the jacobians
-
-    # derive the body to nav direction transformation matrix
-    Tbn = sf.Matrix([[sf.cos(state[State.yaw]) , -sf.sin(state[State.yaw])],
-                [sf.sin(state[State.yaw]) ,  sf.cos(state[State.yaw])]])
-
-    # attitude update equation
-    yaw_new = state[State.yaw] + d_ang
-
-    # velocity update equations
-    v_new = sf.V2(state[State.vx], state[State.vy]) + Tbn * d_vel
-
-    # Define vector of process equations
-    state_new = VState.block_matrix([[v_new], [sf.V1(yaw_new)]])
+    state = State.from_tangent(state)
+    d_ang = sf.V1(d_ang) # cast to vector to gain group properties (e.g.: to_tangent)
+
+    state_error = Values(
+        vel = sf.V2.symbolic("delta_vel"),
+        yaw = sf.V1.symbolic("delta_yaw")
+    )
+
+    # True state kinematics
+    state_t = Values(
+        vel = state["vel"] + state_error["vel"],
+        R = state["R"] * rot2_small_angle(state_error["yaw"])
+    )
+
+    noise = Values(
+        d_vel = sf.V2.symbolic("a_n"),
+        d_ang = sf.V1.symbolic("w_n"),
+    )
+
+    input_t = Values(
+        d_vel = d_vel - noise["d_vel"],
+        d_ang = d_ang - noise["d_ang"]
+    )
+
+    state_t_pred = Values(
+        vel = state_t["vel"] + state_t["R"] * input_t["d_vel"],
+        R = state_t["R"] * rot2_small_angle(input_t["d_ang"])
+    )
+
+    # Nominal state kinematics
+    state_pred = Values(
+        vel = state["vel"] + state["R"] * d_vel,
+        R = state["R"] * rot2_small_angle(d_ang)
+    )
+
+    # Error state kinematics
+    delta_rot = (state_pred["R"].inverse() * state_t_pred["R"])
+    state_error_pred = Values(
+        vel = state_t_pred["vel"] - state_pred["vel"],
+        yaw = sf.simplify(delta_rot.z.imag) # small angle appriximation; use simplify to cancel R.T*R
+    )
+
+    zero_state_error = {state_error[key]: state_error[key].zero() for key in state_error.keys()}
+    zero_noise = {noise[key]: noise[key].zero() for key in noise.keys()}
 
     # Calculate state transition matrix
-    F = state_new.jacobian(state)
+    F = VTangent(state_error_pred.to_storage()).jacobian(state_error).subs(zero_state_error).subs(zero_noise)
 
     # Derive the covariance prediction equations
     # Error growth in the inertial solution is assumed to be driven by 'noise' in the delta angles and
     # velocities, after bias effects have been removed.
 
-    # derive the control(disturbance) influence matrix from IMU noise to state noise
-    G = state_new.jacobian(sf.V3.block_matrix([[d_vel], [sf.V1(d_ang)]]))
+    # derive the control(disturbance) influence matrix from IMU noise to error-state noise
+    G = VTangent(state_error_pred.to_storage()).jacobian(noise).subs(zero_state_error).subs(zero_noise)
 
     # derive the state error matrix
     var_u = sf.Matrix.diag([d_vel_var, d_vel_var, d_ang_var])
@@ -93,15 +128,15 @@ def yaw_est_predict_covariance(
     # Generate the equations for the upper triangular matrix and the diagonal only
     # Since the matrix is symmetric, the lower triangle does not need to be derived
     # and can simply be copied in the implementation
-    for index in range(State.n_states):
-        for j in range(State.n_states):
+    for index in range(State.tangent_dim()):
+        for j in range(State.tangent_dim()):
             if index > j:
                 P_new[index,j] = 0
 
     return P_new
 
 def yaw_est_compute_measurement_update(
-        P: MState,
+        P: MTangent,
         vel_obs_var: sf.Scalar,
         epsilon : sf.Scalar
 ):
@@ -123,8 +158,8 @@ def yaw_est_compute_measurement_update(
     # Generate the equations for the upper triangular matrix and the diagonal only
     # Since the matrix is symmetric, the lower triangle does not need to be derived
     # and can simply be copied in the implementation
-    for index in range(State.n_states):
-        for j in range(State.n_states):
+    for index in range(State.tangent_dim()):
+        for j in range(State.tangent_dim()):
             if index > j:
                 P_new[index,j] = 0
 

src/modules/ekf2/EKF/python/ekf_derivation/generated/yaw_est_predict_covariance.h

@@ -20,6 +20,7 @@ namespace sym {
  *     P: Matrix33
  *     d_vel: Matrix21
  *     d_vel_var: Scalar
+ *     d_ang: Scalar
  *     d_ang_var: Scalar
  *
  * Outputs:
@@ -29,13 +30,13 @@ template <typename Scalar>
 void YawEstPredictCovariance(const matrix::Matrix<Scalar, 3, 1>& state,
                              const matrix::Matrix<Scalar, 3, 3>& P,
                              const matrix::Matrix<Scalar, 2, 1>& d_vel, const Scalar d_vel_var,
-                             const Scalar d_ang_var,
+                             const Scalar d_ang, const Scalar d_ang_var,
                              matrix::Matrix<Scalar, 3, 3>* const P_new = nullptr) {
-  // Total ops: 33
+  // Total ops: 39
 
   // Input arrays
 
-  // Intermediate terms (7)
+  // Intermediate terms (8)
   const Scalar _tmp0 = std::cos(state(2, 0));
   const Scalar _tmp1 = std::sin(state(2, 0));
   const Scalar _tmp2 = -_tmp0 * d_vel(1, 0) - _tmp1 * d_vel(0, 0);
@@ -44,6 +45,7 @@ void YawEstPredictCovariance(const matrix::Matrix<Scalar, 3, 1>& state,
       std::pow(_tmp0, Scalar(2)) * d_vel_var + std::pow(_tmp1, Scalar(2)) * d_vel_var;
   const Scalar _tmp5 = _tmp0 * d_vel(0, 0) - _tmp1 * d_vel(1, 0);
   const Scalar _tmp6 = P(1, 2) + P(2, 2) * _tmp5;
+  const Scalar _tmp7 = std::pow(d_ang, Scalar(2)) + 1;
 
   // Output terms (1)
   if (P_new != nullptr) {
@@ -55,9 +57,9 @@ void YawEstPredictCovariance(const matrix::Matrix<Scalar, 3, 1>& state,
     _p_new(0, 1) = P(0, 1) + P(2, 1) * _tmp2 + _tmp3 * _tmp5;
     _p_new(1, 1) = P(1, 1) + P(2, 1) * _tmp5 + _tmp4 + _tmp5 * _tmp6;
     _p_new(2, 1) = 0;
-    _p_new(0, 2) = _tmp3;
-    _p_new(1, 2) = _tmp6;
-    _p_new(2, 2) = P(2, 2) + d_ang_var;
+    _p_new(0, 2) = _tmp3 * _tmp7;
+    _p_new(1, 2) = _tmp6 * _tmp7;
+    _p_new(2, 2) = P(2, 2) * std::pow(_tmp7, Scalar(2)) + d_ang_var;
   }
 }  // NOLINT(readability/fn_size)
 

src/modules/ekf2/test/CMakeLists.txt

@@ -57,7 +57,6 @@ px4_add_unit_gtest(SRC test_EKF_terrain_estimator.cpp LINKLIBS ecl_EKF ecl_senso
 px4_add_unit_gtest(SRC test_EKF_utils.cpp LINKLIBS ecl_EKF ecl_sensor_sim)
 px4_add_unit_gtest(SRC test_EKF_withReplayData.cpp LINKLIBS ecl_EKF ecl_sensor_sim)
 px4_add_unit_gtest(SRC test_EKF_yaw_estimator.cpp LINKLIBS ecl_EKF ecl_sensor_sim ecl_test_helper)
-px4_add_unit_gtest(SRC test_EKF_yaw_estimator_generated.cpp LINKLIBS ecl_EKF ecl_test_helper)
 px4_add_unit_gtest(SRC test_EKF_yaw_fusion_generated.cpp LINKLIBS ecl_EKF ecl_test_helper)
 px4_add_unit_gtest(SRC test_SensorRangeFinder.cpp LINKLIBS ecl_EKF ecl_sensor_sim)
 px4_add_unit_gtest(SRC test_EKF_drag_fusion.cpp LINKLIBS ecl_EKF ecl_sensor_sim)