Modeling and Identification of Probability Distributions in Robot Kinematics for Increased Precision in Manipulation Tasks

Abstract

In scenarios where teleoperation is not possible, manipulation actions need to be reliable. This is especially true for robotic space exploration, where planetary exploration rovers are expected to handle and transport equipment in future missions, but commonly employ lightweight robotic arms which have traveled thousands of kilometers through space and need to perform in environments with large temperature changes. To ensure reliable and accurate manipulation despite these challenges, this thesis presents a calibration procedure for a probabilistic robot forward kinematics model based on the Product-of-Exponentials (POE) formula. We model a kinematic chain as perturbed by both systematic effects such as geometric and compliant offsets as well as effects assumed to be random, such as gear backlash, material deformation or component wear. Our calibration then identifies the systematic offsets as the distributions’ means using traditional forward kinematics calibration, while the covariance matrices are identified using a novel maximum-likelihood-based algorithm. In addition, we present a novel method to evaluate manipulation approach configurations with regard to their uncertainty. This enables choosing configurations which minimize task-relevant uncertainty at the end-effector based on the identified uncertainties within the kinematic chain. Based on data of the Kinova Jaco2 manipulator of German Aerospace Center (Ger. Deutsches Zentrum für Luft- und Raumfahrt e.V.) (DLR)’s exploration rover Lightweight Rover Unit (LRU), we then show that this combined probabilistic calibration improves the forward kinematics accuracy by 60% to 73% and that the calibrated covariances can describe end-effector pose deviations better than covariances set from experience based on a likelihood-ratio test. In addition, we validate our configuration evaluation in a real-world example: in an experiment where the LRU 2 was tasked with picking up payload boxes from a lander, using low uncertainty approach configurations reduces the task-relevant error by roughly 30% compared to higher uncertainty configurations. Therefore, our work can be seen as proof-of-concept on how a probabilistic robot model can be used to improve the reliability of manipulation actions.