In L18.20 EKF exercise, sigma is not properly initialized. `predict` method is not properly implemented

[xcarnd]

In this exercise, the initial value of sigma is initialized as:

sigma_0 = np.matmul(np.identity(3), >np.array([angle_error,velocity_sigma,position_sigma]))

which results in a 1x3 vector. But sigma is the covariance matrix so it shall be a 3x3 matrix.

The implementation for predict is also incorrect:

    def predict(self, u):

        previous_covariance = self.sigma
        mu_bar = self.g(u)
        G_now  = self.g_prime()
        sigma_bar = np.matmul(G_now,np.matmul(previous_covariance,np.transpose(G_now))) + self.r_t

        self.mu_bar  = mu_bar 
        self.sigma_bar = sigma_bar

        return mu_bar, sigma_bar

sigma_bar is added with self.r_t, which is the sensor noise matrix (It is correct in the solution notebook btw).

Similar problem for update (added with self.q_t instead of self.r_t in the student notebook. correct in the solution notebook).

[dyfc]

Fixed sigma_0 issue by making the following changes:

1. In EKF-Student.py

   sigma_0 = np.identity(3) * np.array([angle_error,velocity_sigma,position_sigma])

2. In StateSpaceDisplayFor2D.py

   state_space_display_predict():
   ...
   initial_state_space=multivariate_normal([y,y_dot], [[sigma_0[2,2],0.0],[0.0,sigma_0[1,1]]])
   ...

[annerajb]

Can confirm had to change q_t and r_t

[WillieMaddox]

I came here to submit a bug report and came across this post. I just wasted an hour debugging the same issues as those listed above but it looks like these issues were resolved 7 months ago. Please make sure the changes are getting pushed to current workspaces. Thanks.