1
$\begingroup$

Consider a nonlinear system $x(k+1)=f(x(k),u(k))$, where $x(k)\in\mathbb{R}^{n}$ is the state, $u(k)\in\mathbb{R}^m$ is the control input. Here $u(k)$ is normally distributed RV with mean $\mu_u(k)$ and variance $\Sigma_u(k)$. I want to find the distribution of $x(k)$ starting from some deterministic $x(0)$ for some horizon $K$. Then can the ordinary EKF update equation be used for the prediction in the below way?

\begin{align*} \mu_x(k+1)&=f(\mu_x(k),\mu_u(k)),\\ \Sigma_x(k+1)&=\nabla_x f(\mu_x(k),\mu_u(k))\Sigma_x(k)\nabla_x f(\mu_x(k),\mu_u(k))^\top+\nabla_u f(\mu_x(k),\mu_u(k))\Sigma_u(k)\nabla_u f(\mu_x(k),\mu_u(k))^\top \end{align*}

| improve this question | |
$\endgroup$
1
$\begingroup$

Yes. This approach is commonly used when IMUs are present because they measure the rate of state change with some uncertainty.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Is there no need to add the term $2\nabla_x f(\mu_x(k),\mu_u(k))\Sigma_{xu}(k)\nabla_u f(\mu_x(k),\mu_u(k))^\top$? $\endgroup$ – Astghik Hakobyan Jul 8 at 4:14
  • 1
    $\begingroup$ Usually $\Sigma_{xu}$ is zero. You'll have to include it if it's not. $\endgroup$ – holmeski Jul 8 at 13:27
  • 1
    $\begingroup$ It being non zero would mean the random control input is correlated with your state. $\endgroup$ – holmeski Jul 8 at 13:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.