Honing the Craft

Introduction

Construction of the groundbreaking Lucas Museum of Narrative Art (LMNA) required the manufacturing of more than 1270 unique, massive, three-dimensionally articulating façade panels and represents a high-water mark for the execution of a mass-customized building cladding system. The project provided an opportunity to advance the craft of digital fabrication for façade manufacturing, but would require significant process development of digital workflow as well as the harnessing of novel technologies such as large-scale industrial robotics and rapid laser scanning to ensure on-schedule fulfillment of the exacting visual and dimensional quality standards for all the façade components.

Free-form, three-dimensional modeling software packages are now ubiquitous within the architecture and design industries, and have been steadily entrenching themselves within the design disciplines for decades. These programs allow architectural designers to intuitively manipulate building envelope form to fantastic effect with the Lucas Museum being a prime example. But to date, physically constructing these natively digital creations has typically been accomplished with more traditional construction materials such as timber, concrete, metals, or glass.

However, one consequence of the organic, fluid design language employed in the building’s exterior form is the inherent irrationality of the surfaces: they do not lend themselves to being reduced to simpler - and more importantly, repeatable - geometric figures. This obstacle, coupled with the architect’s desire to minimize the quantity of panel-to-panel joints yielded panels of truly enormous size and variability, presenting a daunting manufacturability challenge, one not possible to resolve with more conventional materials of construction either due to weight constraints or the inability to form the surfaces economically in a given material.

Composite materials such as FRP (fiber-reinforced polymers) presented a solution that was sufficiently lightweight and viably cost-effective. Perhaps most importantly, as a molded material, FRP possessed the freedom of form necessary to resolve the complex surfacing of the building’s exterior envelope through a variety of mold-making fabrication processes.

As if the challenges of scale and geometric complexity were not enough, the question of how to achieve the architect’s specified surface finish compounded these issues, exponentially complicating the manufacturing challenge. At the project’s core, a question emerged: how to cost-efficiently produce a uniform finish in a novel material on 1273 unique geometries encompassing nearly 200,000 square feet of surface area?

This paper explores the particular challenges encountered & overcome in the course of completing the FRP façade cladding for the Lucas Museum of Narrative Art, and will focus on specific issues related to an agile and flexible 3D modeling framework and robotic programming for manufacturing.

Background & Challenges

Geometric Variation & Robotic Reach

First, we must contextualize the manufacturing challenge presented by the physical scale and scope of geometric variation of LMNA’s cladding panels and recognize its impacts on any potential finishing process. A given panel assembly’s minimum panel face bounding box dimensions in local panel coordinates (bbox L, bbox W, and bbox H) are defined for illustration purposes in the two panels below (Figures 1 and 2). The full breadth of panel variation is shown in a scatter plot of these criteria (Figure 3) with the X-axis showing an overall panel length variation from 6-50 feet, the Y-axis indicating panel width variation from 3-12 feet, and the radii of each dot representing the panel depth variation from 0.5 - 5.25 feet. Colors in this scatter plot represent unique panel typologies including panels with secondary or integral gutter elements, panels with reveals, panels with return features, etc., each with their own special considerations to accommodate their individual finishing conditions. This is seen perhaps most concretely with a series of images of actual project panels loaded for finishing operations in the robotic work cell (Figure 4).

One of the fundamental challenges this panel landscape presents is with regard to the robot’s capacity to engage the entirety of the panel surface. As seen below here (Figure 5), the hatched area represents the area the robot is capable of reaching. Note that a portion of the particular panel shown here sits outside this “operable bubble”, indicating that the robot would be unable to reach and hone this section of the panel’s surface.

Unfortunately, this diagram, albeit useful, represents a greatly simplified view of robotic positioning capacity and motion control. It is a necessary but insufficient condition, as it still fails to account for tool orientation which, in the end effector configuration selected, must always remain normal to the panel surface during honing operations.

Owing to the great variety of panel sizes and shapes, developing kinematic solutions was initially very manual, cumbersome, and much more subject to programmer intuition than to a set of well-defined objective criteria.

Required Material Removal

Additionally, the mold surface texture artifact on raw, demolded panels (Figure 6) vs. the desired smooth, matte finish specification (Figure 7) posed particular challenges throughout the fabrication process that ultimately needed to be resolved in the robotic finishing process.

As a composites industry standard, gelcoat is typically sprayed in thicknesses ranging from 0.25 mm to 1 mm (10 to 40 mils), with most cases involving a nominal thickness of around 0.5 mm (20 mils). For LMNA, owing to the surface texture inherent in the molding strategy and the process controls available at the scale of the panels for the project, gelcoat needed to be sprayed between 80-120 mils to account for the minimum material removal characteristic required to account for mold texture, mold seams, etc. (Figure 8).

This degree of material removal ultimately required rapid process development to move away from the tooling initially planned for the process (Figure 9), and toward the implementation of much more aggressive cutting tools (Figure 10) in order to avoid excessive repeated honing cycles. But this left little room for errors in panel processing - both digital and physical - and presented a high-risk production scenario: excessive honing leading to FRP exposure (Figure 11) typically results in a scrapped part as the damage cannot be repaired without compromising the finish quality of the final panel.

Surface Conformance

As panels of increasing size and complexity were brought to the robotic cell for honing, the results of the honing process started to become inconsistent. Only upon closer inspection via laser tracker, could it be determined that panels were not maintaining dimensional conformance between their prone and upright orientations (Figure 12). Put another way, panels no longer reflected their nominal CAD geometry with the precision and accuracy needed for quality-assured, automated honing.

For certain panels, the surface deviations seen between the panel in these two configurations were so large that they exceeded the full stroke capacity of the end effector’s applied force compliance device, rendering the panel incapable of being honed without time-consuming tracker scans performed by an operator inside the cell, and highly-skilled digital surface reconstruction, rendering robots inoperable until the scan and processing could be completed - approximately 120-180 minutes per panel depending on size and operator experience. And just as importantly, panels entered the robot cell with little predictability as to which panels would require such lengthy, disruptive treatment until significant time had already been spent programming for it. With the robotic work cell already a bottlenecked resource, this condition was unsustainable and threatened the robotic honing process as a scalable, project-wide finishing solution.

Process

Panel Placement & Kinematics Solutions

In order to position panels in the work cell such that they were capable of being honed successfully, the team needed to produce unique placement scheme diagrams that defined several criteria: work cell position, panel orientation (upright vs. inverted), vertical clearances, perpendicular offset from the wall fixture at 4 pre-defined points, etc. (Figure 13). In some cases, multiple placements, and consequently diagrams, were necessary due to the inability to hone a panel in its entirety from a single orientation.

For this task, the team developed an end-to-end, templated digital pipeline using Rhinoceros/Grasshopper 3D (McNeel/Rutten), PowerMILL (version 2019-2024, Autodesk) and RoboDK (version 5.2.2, RoboDK, 2021) that ingested the CAD fabrication model geometry of both the FRP panel and its corresponding steel frame, situated it within the virtual robot cell environment, produced the required documentation, and passed the surface to the robotic programming environment, ultimately simulating the honing program to produce a verified kinematics solution. These outputs from the honing digital pipeline all form individual components of the assembled panel traveler document submitted to the shop floor for panel production.

As for solving kinematics, as the team gained experience honing panels of various general geometric “families”, patterns of kinematic solutions emerged, leading to generalized robotic joint priorities that yielded much more consistent programming results. These joint priorities were compiled into tables, and iterated through by custom Python scripts within the RoboDK environment to rapidly develop valid honing programs.

Scanning & Location Verification

To resolve the lack of in-cell panel conformance to the nominal CAD geometry, we determined that developing accurate representations of the panel surface as it is presented to the robot for honing is the ideal case as it ensures the toolpaths written conform to the exact panel surface as-rigged in the robot cell (Figure 15). Then, it was simply a matter of our existing SMR-based (spherical mirror-reflector) laser tracker tools being unfit for this task. We turned to laser scanning instead, eventually settling on a Leica RTC360 unit paired with Leica’s Cyclone Register360 software, allowing operators with less than hour of specialized training to scan panels with guaranteed dimensional fidelity in under 20 minutes per panel (Figures 16, 17).

The team supplemented this rapid reality capture workflow with its own Rhino / Grasshopper → PowerMILL → RoboDK pipeline (Figure 18) for processing the scan data, developing all honing toolpaths, and exporting verified robotic programs.

The pipeline considers and accounts for various panel conditions, balancing the need for consistent quantity of material removal with a preference to keep the tool engaged with the panel face throughout the honing program for minimal disengagement / reengagement with the surface to avoid unnecessary gouging of the panel face, resulting in two divergent machining strategies, referred to as parallel and tween strategies (Figures 19, 20, respectively).

For indexing the location of the part within the robot cell, operators also collect robotic joint position values at drilled target points (Figure 21) which are also captured in the scan data, allowing a reference frame to be generated corresponding to the output workplane for all toolpaths generated downstream for the panel in question.

As a verification procedure, in addition to all the honing programs, at least one extra program is exported for each panel referred to as the ‘location’ program. This is a non-contact program, set off the panel face and inset from the panel edge by 1”. This program is developed with a custom, internally developed range-finder tool using an inexpensive time-of-flight sensor and a battery-powered Microsoft Particle Argon board, all contained in a compact 3d-printed housing and mounted on a standard toolholder (Figure 22). This sensor pulses a laser at a regular interval around the perimeter of the panel, reading out its distance from the panel using the length of the profiling toolpath to trigger the total number of pulses.

This data is read into a templated spreadsheet (Figure 23) that calculates the average, min, max, and median distance from the panel, and most importantly, the total range. As long as the range is safely below the throw of the applied force controller on the end effector, the panel can be safely honed by re-centering the tools for the programs to the middle of the range. This task is easily accomplished within 5-10 minutes of receiving the location data and provided the tool sits normal to the surface, ensures safe and accurate honing on every panel.

End-to-end, inactive time on a typical panel has been reduced to less than 45 minutes from the time it is secured in its position in the cell to when the robotic operators have a complete set of honing programs in-hand to execute. At peak production capacity, a small programming team is capable of writing the equivalent of 15 linear miles of finishing toolpath per day, all of which is unique and never to be repeated or re-used on the project.

Conclusion and Future Work

Process development of the robotic honing digital workflow was vital to the production of the freeform exterior façade panels of the Lucas Museum of Narrative Art, a watershed project from any standpoint, but with particular regard to building product materials and manufacturing methods. Achieving these levels of fit and finish at this scale and schedule were only possible with robust yet flexible automation strategies paired with an understanding of the complexities presented by the geometries and materials.

In the development of these workflows, we found it necessary to productively subvert the assumption of the inherited CAD geometry as the single 'source-of-truth' for robotic finishing processes. Only by leaving behind rigid assumptions of geometric fidelity to this digital precursor was the robotic finishing process able to be scaled to meet the demands of the project. This suggests potential territory for further exploration in areas where tolerances to source surface CAD data are unclear, but can be captured in a 'just-in-time' paradigm - a cornerstone principle of lean manufacturing - through rapid reality capture tools which will only become more economical, user-friendly, and ubiquitous as time progresses.

Future projects would seek to further the reach of automation for near-repetitive elements or "families", extend scanning to provide panel location within the cell environment as well as global shape, and may require the existing digital pipeline and physical finishing process to be adapted to accommodate more intricate localized curvature, finer levels of surface patterning and texture, and dynamic control of applied force based on local surface conditions.

Lucas Museum of Narrative Art, Los Angeles, CA

Client: Lucas Museum of Narrative Art

Architect: MAD Architects

Executive Architect: Stantec

Panel FRP Fabrication: NCMS

Steel System: Martin Bros.

General Contractor: Hathaway Dinwiddie Construction Company

Scott Bader Company Limited. Crystic Polyester Handbook, Wollaston, Wellingborough, Northamptonshire. (2005).

Yuhazri, M. & Husin, Mohd Amirhafizan & Nilson, G. & Kamarul, A. & Nirmal, Umar. (2018). Effects of gelcoat thickness on mechanical properties of woven glass/polyester laminated composite. JOURNAL OF MECHANICAL ENGINEERING AND SCIENCES. 12. 3370-3386. 10.15282/jmes.12.1.2018.7.0301.

Robert McNeel & Associates (2020) Rhino 3D (Version 7.0) [Computer program]. Available at: https://www.rhino3d.com/downlo... (Accessed 21 December 2020).

Autodesk, Inc. (2018-2023) PowerMILL (Version 2019-2024) [Computer program]. Available at: https://manage.autodesk.com/pr... (Accessed 7 October 2019 - 22 May 2023).

RoboDK, Inc. (2021) RoboDK (Version 5.2.2) [Computer program]. Available at: https://robodk.com/download (Accessed 23 March 2021).

Leica Geosystems (2021) Cyclone Register 360 (Version 2021.1.1) [Computer program]. Available at: https://connect.leica-geosyste... (Accessed 28 June 2021).