Core Concepts
Deriving analytical backpropagation equations for EKF covariance gradients enables perception-aware optimal motion planning.
Abstract
The article introduces novel backpropagation analytical equations for deriving the gradients of an Extended Kalman Filter's final covariance matrix. It discusses perception-aware trajectory optimization, simulations validating the approach, and real-world experiments on an off-road vehicle. The method showcases improvements in localization accuracy and execution time.
Stats
"Deriving novel analytical backpropagation equations for the gradient of the covariance of an EKF with respect to all inputs of the filter."
"Applying the technique to derive a computationally efficient perception-aware method."
"The computation time for the gradient of the loss with respect to control inputs using backpropagation is 0.19 ± 0.07s."
Quotes
"In this paper, we fill a gap by providing closed-form analytical expressions for the derivatives of any smooth function of the final covariance matrix PN of an EKF, w.r.t. all previous control inputs."