Date of Award

Spring 1-1-2014

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Nikolaus Correll

Second Advisor

Sriram Sankaranarayanan

Third Advisor

Tom Yeh

Fourth Advisor

Eric Frew

Fifth Advisor

Daniel Scheeres

Abstract

Assembly robots have been in operation in industry for decades, predictably repeating the same precise motions in closed workspaces to assemble products cheaply and in mass quantities. However, in the field, robotic assembly has seen only spurts of progress, and no short-term feasible applications. NASA and the space industry desire robotic construction methods to remove the upper limit on size. Space telescopes are highly desired, but require structural precision on the order of microns. Previous approaches were ruled out because the precisely machined components were expensive, heavy, and prone to failure.

The recent advent of cheap robotic swarms has revived interest in academia, but most research requires self-correcting, interlocking components, instead of commodity materials.

In this thesis, I describe the Intelligent Precision Jigging paradigm, a solution to the problem of practical robotic assembly, with application to precision truss assembly. Intelligent Precision Jigging Robots (IPJRs) are robots that work in groups of three to incrementally assemble a structure. They set and hold distances with high precision, enabling coarse external manipulators to weld the commodity parts together and perform other tasks.

To maximize the utility of the IPJR paradigm to the fullest extent, I present algorithms for finding near-optimal assembly sequences and for implementing Simultaneous Localization and Mapping (SLAM) to maintain an estimate of the assembly process through the accumulation of local strut length measurements. I define a model of truss assembly probability and a minimizing metric based on the covariance trace. I show that structure error grows cubically with node count.

I present a three-step approach for generating near-optimal assembly sequences; commencing assembly on a central location of the structure, greedily assembling to minimize the covariance trace, and performing a local search on the space of sequences to swap steps until a local minimum is found. I show that this method consistently generates more precise sequences than any process alone.

I then simulate the SLAM method with four different estimators commonly used in for SLAM; a least linear squares approach, the Extended Kalman Filter, the Unscented Kalman Filter, and the Maximum Likelihood Estimator. I show that when nonlinearity in the assembly process is dominant, the Maximum Likelihood Estimator is better than the other estimators, but for space telescopes with precision requirements, all four are functionally equivalent. I also show that when SLAM is used, the difference in covariance trace between sequences is reduced, reducing the need for finding globally optimal sequences. SLAM also mitigates the growth of structure error.

Finally, I present the results of physical assembly trials on a telescope truss made of aluminum tubes, assembled by three IPJRs using two methods: an open loop approach, and an MLE-SLAM approach. I show that the MLE-SLAM assembly algorithm works even when the physical trials included unmodeled processes such as deformation under gravity, outperforming the open loop algorithm.

Share

COinS