Visual Pursuit Control with Target Motion Learning via Gaussian Process

Abstract

In this paper, we propose an observer-based visual pursuit control integrating 3-dimensional target motion learning by Gaussian Process Regression (GPR). We consider a situation where a visual sensor equipped rigid body pursuits a target rigid body whose velocity is unknown, but dependent on the target's pose. We estimate the pose from visual information and propose a Gaussian Process (GP) model to predict the target velocity from the pose estimate. We analyze stability of the proposed control by showing that estimation and control errors are ultimately bounded with high probability. Finally, simulations illustrate the performance of the proposed control schemes even if the visual measurement is corrupted by noise.

Gaussian process-based visual pursuit control with unknown target motion learning in three dimensions

“…This is a wider class of the body velocity than that assumed in [15] where the class of body velocity is assumed to be dependent just on the position p wo , namely, V b wo : R 3 → R 6 . This paper uses a GP model to identify the unknown body velocity V b wo as a function of the poseǧ wo , and use it for visual pursuit control later in Section 4.…”

“…The goal is now to combine the VMO with the GP model in feed-forward fashion from the previous Section 3 for a visual pursuit control scenario. In contrast to the author's previous publication [15], the full pose is incorporated for the control scheme. Further, differently to Section 3, it is not assumed thatǧ wo ≈ǧ wo since it is too risky in a control setting to use large gains that might amplify noise severely.…”

“…As a remark, the controller is similar to [15], but rather than using only the estimate target position as an input for the GP model, both estimates of position and orientation are used in this approach, since a wider class of the target velocity is considered as given by Equation (5). Furthermore, let the controller K ǧ wo be of the form…”

“…By using the mean function of a GP model in a feed-forward manner, Beckers et al [13] aims to eliminate unmodelled dynamics and further provide a method to adjust the error feedback gains based on the GP variance. The author's recent publication [15] extends the result in [9] with the technique from [13] by integrating the learned GP target motion model into the visual pursuit control scheme to guarantee stability with high probability. In Ref.…”

“…In Ref. [15], a GP model learns the unknown body velocity as a function of the target's position for an observer-based control, assuming that the target is moving influenced by the environment, such as terrains and obstacles. However, since the target motion consists of both translation and rotation, the proposed visual pursuit control scheme is limited to a special class of target motions.…”