Precise assembly of 3D truss structures using MLE-based error prediction and correction

We describe a method to construct precise truss structures from non-precise commodity parts. Trusses with precision in the order of micrometers, such as the truss of a space telescope, have previously been built using precisely machined truss connection systems. This approach is expensive, heavy, and prone to failure, e.g. when a single element is lost. In the past, we have proposed a novel concept in which non-precise commodity parts can be aligned using intelligent precision jigging robots and then welded in place. Even when using highly precise sensors and actuators, this approach can still lead to errors due to thermal expansion and structural deformation. In this paper, we describe and evaluate algorithms for generating truss assembly sequences that reduce the expected error by (1) using a heuristic to generate build sequences that reduce the expected variance, and (2) during assembly, estimating the structure's pose using maximum likelihood estimation that combines local measurements by different intelligent precision jigging robots, improves this estimate during loop closures in the construction process, and uses this estimate to correct for errors during construction. We show through simulation and physical experiment that this combined approach reduces assembly error, enabling precision construction with commodity materials. While the model herein is based on truss structures, the proposed methods generalize to a larger class of incremental assembly problems, which exhibit continuous rather than discrete errors.