Model estimation and control of compliant contact normal force

Abstract

This paper proposes a method to realize desired contact normal forces between humanoids and their compliant environment. By using contact models, desired contact forces are converted to desired deformations of compliant surfaces. To achieve desired forces, deformations are controlled by controlling the contact point positions. Parameters of contact models are assumed to be known or estimated using the approach described in this paper. The proposed methods for estimating the contact parameters and controlling the contact normal force are implemented on a LWR KUKA IV arm. To verify both methods, experiments are performed with the KUKA arm while its end-effector is in contact with two different soft objects.

Robot-Safe Impacts with Soft Contacts Based on Learned Deformations

“…In recent work, [11] proposed a method for soft contact model identification. The approach in [12] combines estimation and control of the material's contact normal stiffness in a single framework.…”

“…Solving (11) multiple times for different rays ω, we collect a list of extreme vertices and halfplanes, representing a convex polytope. Refer to [30] for an efficient algorithm to select useful search directions (rays).…”

“…The soft material properties need to be learned (and updated) in a data-driven way unless precise knowledge is apriori available [11]. In order to collect the required data, the robot has to penetrate the deformable contact surfaces.…”

“…Given a large amount of training data, (17) can only be applied for off-line learning because of the time-consuming matrix inversion. We here extend [11] by employing the recursive least-squares algorithm [32, page 196] that is suitable for model updates in real-time. Given a new data point consisting of s i and f i , we update in each control cycle i > 0 the intermediate terms (denoted in subscripts)…”

“…For reproducibility, we ordered a similar dice as chosen by [11] and also selected the vertical axis as the direction of motion, typical for bin-picking scenarios. The soft dice weights 0.…”