MIMO ILC for Precision SEA robots using Input-weighted Complex-Kernel Regression
Authors:
Leon Yan,
Nathan Banka,
Parker Owan,
Walter Tony Piaskowy,
Joseph Garbini,
Santosh Devasia
Abstract:
This work improves the positioning precision of lightweight robots with series elastic actuators (SEAs). Lightweight SEA robots, along with low-impedance control, can maneuver without causing damage in uncertain, confined spaces such as inside an aircraft wing during aircraft assembly. Nevertheless, substantial modeling uncertainties in SEA robots reduce the precision achieved by model-based appro…
▽ More
This work improves the positioning precision of lightweight robots with series elastic actuators (SEAs). Lightweight SEA robots, along with low-impedance control, can maneuver without causing damage in uncertain, confined spaces such as inside an aircraft wing during aircraft assembly. Nevertheless, substantial modeling uncertainties in SEA robots reduce the precision achieved by model-based approaches such as inversion-based feedforward. Therefore, this article improves the precision of SEA robots around specified operating points, through a multi-input multi-output (MIMO), iterative learning control (ILC) approach. The main contributions of this article are to (i) introduce an input-weighted complex kernel to estimate local MIMO models using complex Gaussian process regression (c-GPR) (ii) develop Geršgorin-theorem-based conditions on the iteration gains for ensuring ILC convergence to precision within noise-related limits, even with errors in the estimated model; and (iii) demonstrate precision positioning with an experimental SEA robot. Comparative experimental results, with and without ILC, show around 90% improvement in the positioning precision (close to the repeatability limit of the robot) and a 10-times increase in the SEA robot's operating speed with the use of the MIMO ILC.
△ Less
Submitted 23 October, 2020; v1 submitted 9 October, 2020;
originally announced October 2020.
Iterative Machine Learning for Precision Trajectory Tracking with Series Elastic Actuators
Authors:
Nathan Banka,
W. Tony Piaskowy,
Joseph Garbini,
Santosh Devasia
Abstract:
When robots operate in unknown environments small errors in postions can lead to large variations in the contact forces, especially with typical high-impedance designs. This can potentially damage the surroundings and/or the robot. Series elastic actuators (SEAs) are a popular way to reduce the output impedance of a robotic arm to improve control authority over the force exerted on the environment…
▽ More
When robots operate in unknown environments small errors in postions can lead to large variations in the contact forces, especially with typical high-impedance designs. This can potentially damage the surroundings and/or the robot. Series elastic actuators (SEAs) are a popular way to reduce the output impedance of a robotic arm to improve control authority over the force exerted on the environment. However this increased control over forces with lower impedance comes at the cost of lower positioning precision and bandwidth. This article examines the use of an iteratively-learned feedforward command to improve position tracking when using SEAs. Over each iteration, the output responses of the system to the quantized inputs are used to estimate a linearized local system models. These estimated models are obtained using a complex-valued Gaussian Process Regression (cGPR) technique and then, used to generate a new feedforward input command based on the previous iteration's error. This article illustrates this iterative machine learning (IML) technique for a two degree of freedom (2-DOF) robotic arm, and demonstrates successful convergence of the IML approach to reduce the tracking error.
△ Less
Submitted 5 October, 2017;
originally announced October 2017.
