Content uploaded by Maria Smigielska
Author content
All content in this area was uploaded by Maria Smigielska on Aug 01, 2018
Content may be subject to copyright.
Application of Machine Learning within the integrative
design and fabrication of robotic rod bending processes.
Maria Smigielska
maria@mariasni.com
Abstract
This paper presents the results of independent research that aims to in-
vestigate the potential and methodology of using Machine Learning
(ML) algorithms for precision control of material deformation and in-
creased geometrical and structural performances in robotic rod bending
technology (RBT). Another focus lies in integrative methods where de-
sign, material properties analysis, structural analysis, optimization and
fabrication of robotically rod bended space-frames are merged into one
coherent data model and allows for bi-directional information flows,
shifting from absolute dimensional architectural descriptions towards
the definition of relational systems. The working methodology thus
combines robotic RBT and ML with integrated fabrication methods as
an alternative to over-specialized and enclosed industrial processes. A
design project for the front desk of a gallery in Paris serves as a proof of
concept of this research and becomes the starting point for future devel-
opments of this methodology.
Keywords: robotic rod bending, machine learning, adaptive processes
2
1 Background
The bending operation is most generally a cold forming process, in which metal is
stressed beyond its yield strength. Until 1991 advancement in rod bending technolo-
gy (RBT) remained insignificant and the process was still operating with traditional
table benders, with a few minor innovations in electronics and safety controls [1]. The
first conceptual schemes of computer integration for the design, fabrication, delivery
and placement of steel reinforcement bars [2] appeared in 1991 and shifted this tech-
nology from a labour-intensive and hazardous one to a safer, faster and almost fully
automated process. Yet, the problem of dependency on human expertise was not
solved, because those top-down systems could not intelligently compensate for varia-
ble springback and other changing conditions. The first truly adaptive system ap-
peared in 2000 [3], when machine learning techniques were applied in this field of
RBT. The first definition of Machine learning (ML)is that it gives computers the
ability to learn without being explicitly programmed [4]. It finds out the highly non-
linear relationships between the given parameters and as such, provides the under-
standing of the analyzed processes by mechanization and enhancement of human
knowledge. The paper mentioned above[3] presents a comparison of the following
prediction methods: 1: multiple linear regression; 2: in-process relaxation and 3: neur-
al networks applied for RBT. All of them result with similar error performance ex-
ceeding +/-2.5°. Such error is acceptable for standard applications of reinforcement
bars, however, it might not be enough for more complex architectural applications
involving metal bars as finished elements or different disciplines that demand higher
precision (medical applications such as orthodontic archwires or spinal fusion surgery
elements). This field requires further investigation.
ML techniques have been widely tested in the metal sheet forming - technology
with similar constraints to rod bending. The Review of Artificial Neural Network
applications [5] from 2012 compares over 60 scientific papers on different approaches
to this topic. It proves that ML techniques can not only achieve zero spring-back an-
gle, but are also used for other applications like process control, process planning,
quality control, finite element simulation, feature recognition, tool design, and cutting
tool selection for sheet metal forming. As such, it reduces complexity, minimizes the
dependency on human expertise and time taken to design the equipment. One of the
recent example projects, titled A Bridge Too Far utilised robotic incremental sheet
forming [6] in2016, employs Neural Networks to gain deeper understanding of the
metal forming process and the resulting imprecision.
Research outcomes of such extent are missing in the field of ML applied to robotic
RBT, even if the applications of these technologies have developed significantly dur-
ing recent years. The project developed by supermanoeuvre for 2012 Venice Architec-
ture Biennale employs robotic rod bending processes and is followed by the research
Adaptive Part Variation [7],where two robotic arms are capable of bending, cutting,
placing, and welding steel rods into a larger assembly with the use of computer vision
systems. Another ongoing research project Mesh Mould Metal [8] is developing robot-
ic wire bending and welding manipulators attached to robotic arm to build formworks
and reinforcements in a single continuous operation.
3
2 Introduction
This paper presents the development of the robotic RBT with incorporation of ML
techniques for the precise control of metal deformation. Another focus lies in integra-
tive methods where design, material properties analysis, structural analysis, optimiza-
tion and fabrication data of robotically rod bended space-frames are merged into one
coherent data model to allow for bi-directional information flows.
A design project for a gallery front desk is taken as a case study.Commissioned by
AA[n+1] Art and Architecture gallery in Paris, the front desk will serve as a piece of
functional furniture (front desk and step to access an elevated office door), as well as
being an exhibited piece in the gallery itself (Fig. 2,6). It is a space frame structure
(5.5m long, 1.4m wide and 0.75 m high) comprising of 6mm cold-drawn S235JR steel
rods, differentiated through a process of non-standard robotic rod bending and welded
on site. The front desk consists of 800 unique elements of varied length (from 0.12 to
1 m), varied number of 2- and 3-dimensional bends (between 2 and 7 bends)and va-
riedangles (2400 unique values between 1° and 144°). The elements are of 341 meters
length and weight 80 kg in total.
In parallel to the front desk project, smaller product design objects are prototyped
to both test new design strategies and verify fabrication processes. In addition, design
and fabrication workshops have been conducted to further test these methods. A four
days students' workshop at ENSA Paris-Malaquais, Digital Knowledge in February
2017 focused on incorporating FEM analysis on two levels: 1: as a design driving
factor- by rationalizing stress lines resulting from specific structural loads/support
conditions and then 2: structurally analyzing it as a second step. Lastly, the prototypes
of 4mm steel rods were fabricated with the robotic RBT (Fig. 14) .
4
3 Methods
3.1 Design Process
The digital workflow for the front desk project is entirely embedded in the Grasshop-
per environment to provide continuity of the information flow between different steps
of design, analysis and production data (Fig. 1). It begins with two parametric guide
curves generating a proto-structural envelope, to be then optimized with a voxel-based
structural analysis (GH Millipede) and evolutionary solver (Galapagos) (Fig. 3). Sec-
ondly, the surface is rationalized with varied morphostructural rod grids and then ana-
lysed with a finite element method (GH Karamba), optimized with another evolution-
ary solver in order to provide a better understanding of the structural conditions (canti-
levering, anchorage, u-shape beam, 3d truss elements) (Fig. 5). For the FEM analysis
the space-frame geometry was simplified to polylines, ignoring the arcs of the bent
parts. Since the structural system differs from the typical space-frame model, a custom
definition for the point welds was created.
The final structural pattern divided the shape into 110 rigid segments, each of
which includes the following elements: left and right principal bars, left and right
diagonal bar of type A and B, which fill the side spaces and two types of zig-zags
pattern that fill the top part of spacing of the frame (Fig. 4).
The digital workflow provides not only a geometrical model and structural calcula-
tions, but also includes the custom made tools written in GH Python for deconstruc-
tion of the curves to prepare data for fabrication and another tool simulating the bend-
ing process in order to predict potential collisions. Such setup stands out from conven-
tional execution planning since there is no need for intermediate conversion to ma-
chine code. The rods are organized with an indexing system and their geometry is
translated as data set containing the information about its angles, orientation (positive-
clockwise, negative-counterclockwise bends) and distances between consecutive
bends to be sent to the text file. Data is fed directly to RobotStudio, where another
custom made tool for bending and translation operations of the robotic arm was writ-
ten (again without any intermediate robot control plugins).
Fig. 1. Data informed design scheme for the front desk project
5
Fig. 2. Functional furniture (front desk and step to access a surelevated office door)
Fig. 3. Generation of the surface by voxel-based structural analysis (Millipede) and evolution-
ary solver
Fig. 4. Final 1:1 mockup and structural pattern of two neighboring segments of the front desk
6
Fig. 5 FROM LEFT TOP. Space-frame deformation (exaggerated representation); FEM
analysis (GH Karamba); structural stiffeners in the step area; welds definition
Fig. 6. The final view of the front desk
7
3.2 Fabrication process (toolheads and software)
In order to simplify the industrial process for RBT, a single robotic arm with no addi-
tional numerically controlled machines was chosen for the fabrication process. Nu-
merous robotic toolheads were iteratively designed, prototyped and tested to find a
simple, yet versatile solution. The rotary station is attached to the robotic arm (as op-
posed to standard CNC machines with table benders) and serves as both bender and
gripper to precisely move the rods along its axis (Fig. 7). The bending operation hap-
pens around the central finger parallel to Z axis of the 6th axis of the robotic arm.
First iteration of the toolhead was dedicated to an ABB4600 robotic arm. The next
iteration of the toolhead is dedicated to an ABB1200 robotic arm and is equipped with
replaceable cylinders for customization of bending radii (Fig. 8) that are installed as
rollers, which reduce the friction between the toolhead and the rod while bending. Due
to the limited weight payload of the ABB1200, the shape of the toolhead is optimized
to its smallest weight.
The rod holding station itself is designed to be paired with different cylinders fit-
ting to rods of different size and topology (round, square, hexagon, from 3 to 6 mm
diameter) as well as with the mechanism preventing the rod from axial rotation under
its own weight (Fig. 8). The station is designed as stand-alone with the single support
to avoid the potential collision with rods while bending (first iteration of the holding
station needed to be attached to the table covered from the side which caused many
rod-table collisions during the process of 3D bending).
Structural system of the front desk differs from typical space-frame setup, based on
nodal joints of several axial elements, however it opens up the opportunity to explore
and recombine different ways of connecting metal elements that preserve its structural
continuity (welding, mechanical splicing, lap splicing). Although the initial explora-
tion of the joinery consists of non standard solutions like structural glues, resin, 3d
printed customized connectors, the final choice led to TIG welding on site (Fig. 9).
Each of 110 segments of the front desk is preassembled and welded separately and
then sequentially welded together to create bigger parts. To provide easy transporta-
tion and assembly on site- the final piece consists of 3 chunks connected with screws.
The bending process is relatively fast- all the 800 elements were bent with one ro-
bot operator within 2 weeks time, but the manual assembly and welding is labour-
intensive and highly error-generating, therefore this is one of the directions of the
research further development (Fig. 13).
8
Fig. 7. Automation of the process of rod bending and axial translation with custom toolhead
attached to the robotic arm.
9
Fig. 8. FROM LEFT TOP: Evolution of the toolhead and customization of equipment: Tool-
head 1.0 for robotic arm ABB 4600; Toolhead 2.0 for the robotic arm ABB 1200 which is
customized according to different bending radii, different rod size and its topology; Overview
of the toolhead elements; stand-alone and single-support rod holding station
Fig. 9. TIG welding, tape, glue, zippers as alternative joinery for preassemble and final assem-
ble phase
10
3.3 Machine learning for material control
In order to produce over 2400 different bending angles, it was necessary to build the
continuous-valued predictive model of the metal behavior. Out of many ML methods-
regression was chosen as it can model a functional relationship between two quantities
(in this case- desired bending angles as X value and its corresponding springback
value Y, calculated as a subtraction of real bent angles from desired bending angles
(Fig. 11)). To create the exhaustive training data, the set of physical bending tests of
various angles between 5° and 200° was done and its springback was measured with
different methods. The first trial involved manual readouts, resulted in imprecisions,
especially for low and high angles (below 30 and over 150°) (Fig. 10 LEFT). Another
method-the averaged photo readouts (averaging several manual readouts from the
photographed samples), resulted in higher precision, allowed to retrieve a n-degree
polynomial (3 degree was finally chosen) predicting the deformations of any desired
angle with the error of +/-1° (Fig. 10 RIGHT). The final dataset consisted of 115
points and was found to be exhaustive. The automation of precise readouts (with com-
puter vision systems), as well as increasing the number of input training data for mul-
tidimensional ML analysis (material properties, environmental data.. ) are some of the
future development goals. ML methods, including polynomial regression and bagging
of polynomial regression were written with an open source scikit-learn library in Py-
thon (Fig. 12).
Fig. 10. LEFT: manual readouts of the bent angles; RIGHT: averaged photo readouts
method
Fig. 11. Diagram representing springback
11
Fig. 12. Polynomial regression and bagging of polynomial regression as examples of machine
learning applied for material deformation control ( X- bent angles, Y- sprinback)
12
Fig. 13. Preassemble, welding, 3 final pieces and its joinery
Fig. 14. Results of the 4days workshop with ENSA Paris-Malaquais, Digital Knowledge, Feb-
ruary 2017
13
14
4 Results and Discussion
The project of the front desk as a demonstrator of robotic RBT has been installed on
site. The ease of preassemble proves the success of the geometric precision of the
bends. It was possible to preserve geometric sophistication without external specialists
and industrial partners by developing simplified, yet versatile solutions for robotic
RBT. The bending process was pursued with one operating person, one robotic arm,
open source software and no additional numerically controlled equipment. This re-
search represents the shift from industrial setup- providing very precise results of spe-
cific application and encapsulated knowledge into the model that is more versatile,
simpler in mechanics, yet of similar precision by incorporating ML to enhance human
expertise.
Moreover, the workshop with 12 students of ENSA Paris-Malaquais, Digital
Knowledge, who had no prior experience to work with the robotic arm, resulted in
prototyping mid size models within just 2 days, which proves the simplicity and ac-
cessibility of the fabrication solutions presented in the paper (Fig. 14).
Nevertheless, in order to grasp the full potential of the methods mentioned above
and fully scale it from prototype to production phase, the following future develop-
ments of the design/fabrication can be drawn:
- Improvement and automation of precise readouts of springback data (by integrat-
ing computer vision systems)
- Clustering multiple robotic processes into one continuous workflow: gripping,
bending, cutting with the emphasis on assembly, which occurs to be the most time
consuming and highly error-generating as manual process
- Improvement of prediction of material elongation, especially for angles higher
than 90 degrees
- Increasing the number of input training data for multidimensional ML analysis by
employing computer vision systems and other data acquisition tools (sensors probing
the outside environment like cameras, laser rangefinders, scanners, temperature sen-
sors or other tools describing internal processes like force or stretch sensors) and test-
ing varied ML techniques (including ANN)
- Attaching the rotary bending station to robotic arm could open up the possibility
to merge bending and positioning in space processes, what could result in significant
reduction of production and assembly time, while keeping the construction precision
on high level
15
5 Final note
Machine learning has the potential to liberate design-to-construction processes of
overly constrained closed industrial systems, in order to finally develop simplified, but
versatile robotic automation for the prototypical building industry. It speculates on
architectural discourse about two notions of precision, control and its various trade-
offs.
Merging back design and making sounds like a very promising scenario, yet it is
not without difficulties to face. The most common upon ML techniques mentioned in
the Review ANN [5], is that experienced persons are required to operate the devel-
oped ANN system. It might seem that ANN does not provide liberty, but only shifts
the control centre from black-boxed industrial knowledge to encapsulated ML exper-
tise. Yet, architecture as a multifarious discipline, has always been embracing differ-
ent fields of knowledge and technologies of its time and that is also the case for now.
It is necessary to gain the digital literacy by embracing ML techniques in order to
achieve full reciprocity between architecture and construction, which in the end allow
a escape from both routine manual work and over constrained industrial systems, to-
wards solutions which are simpler, smarter and more versatile. Such an approach-
stands for creating new possibilities by augmenting human expertise with computa-
tional data, knowledge acquisition and knowledge inference within design, fabrication
and assembly processes.
16
Acknowledgements
The author would like to thank the AREA Institute (Applied Research & Entrepreneurship for
Architecture), Paris
and ABB Robotics France for providing the access to the robotic facility and for the support
during the design and production processes of the front desk.
References
1. KRB Machinery Co.,: DMI standard bender operations and maintenance instruction man-
ual, York, Pa (1993).
2. Miltenberger M. A., Bernold, L. E. : CAD-integrated rebar bending. Proc., 7th Conf.,
Computing in Civ. Engrg. and Symp. on Data Bases, L. F. Cohn and W. Rasdorf, eds, pp.
819–828. ASCE, New York (1991).
3. Dunston Ph.S., Bernold L.E. : Adaptive Control For Safe And Quality Rebar Fabrication,
pp. 122-129. J. Constr. Eng. Manage., 126(2) (2000).
4. Samuel A. L.: "Some Studies in Machine Learning Using the Game of Checkers". IBM
Journal of Research and Development. (1959).
5. Kashid S., Kumar Sh. : Applications of Artificial Neural Network to Sheet Metal Work -
A Review, pp. 168-176. American Journal of Intelligent Systems, 2(7) (2012).
6. Nicholas P., Zwierzycki M., Stasiuk D., Norgaard E., Leinweber S., Thomsen M, 2016:
"Adaptive Meshing for Bi-directional Information Flows: A Multi-Scale Approach to Inte-
grating Feedback between Design, Simulation, and Fabrication", pp.260-273. Advances in
Architectural Geometry (2016).
7. Vasey L., Maxwell I., Pigram D. : A Near Real-Time Approach to Construction Toler-
ances, Adaptive Part Variation, pp. 291-304, Robotic Fabrication in Architecture, Art and
Design 2014 (2014).
8. Hack N., Lauer W., Gramazio F., Kohler M. : Mesh Mould: Robotically Fabricated Metal
Meshes As Concrete Formwork And Reinforcement, FERRO-11: Proceedings of the 11th
International Symposium on Ferrocement and 3rd ICTRC International Conference on
Textile Reinforced Concrete.