
Content
p. 1-4
adaptive pneumatic structure



evolving soft body
parametric roof

min. surface grid shell digital clay

p. 5-7

p. 1-4
adaptive pneumatic structure
evolving soft body
parametric roof
min. surface grid shell digital clay
p. 5-7
explore possible techniques of transferring making knowledge in pottery throwing to machine as a potential aid for design fabrication
Introduction
type Master’s Thesis time 4 months | 2022 keywords simulation, supervised learning
Craftsman’s knowledge in working with medium has a tacit nature, but for the computer-controlled tools this knowledge must be presented or translated into explicit information for fabrication. This project consider an alternate way of communicating design intent to machines without explicitly stating fabrication steps by considering one of the oldest and most tactile making process of pottery throwing. And attempt at transferring the fabrication knowledge into a digital system, by training a Machine Learning (ML) model to learn the causal effect of applying pressure on the final form of pots.
Objective
Objective 1. To develop a simulation of pottery throwing wheel with an entity that approximates the hands applying pressure on clay.
Objective 2. A systematic approach of building a Machine Learning Model by controlling type of model, model hyperparameter and training parameter that can best learn the causal relation.
Objective 3. Qualitatively and quantitatively compare the predicted results for the various models to compare the variance from the mean.
Of the several techniques available for shaping clay, this study focuses on shaping a piece using a potter’s wheel. This process is called wheel throwing or throwing. Shaping a pottery piece requires a series of procedures shown below.
It starts by centering the clay on the wheel. Followed by creating an opening and setting its base depth to create the floor, followed by pulling the walls ups to form a uniform cylinder with equal wall thickness (referred to as the pulled pot). Then finally the shape is given to the pot by pressing in or out at various sections of it (referred to as the formed pot). The last step is to trim and separate the pottery from the wheel . Its the process of forming of the pulled pot (3rd image ) to the formed pot that is focused on (4th image)
Simulation
For a machine to formulate an understanding of the underlying effect of the tool on the clay form, the data from the physical phenomenon must first be represented numerically and pre-processed in certain ways which is machine readable.
So a simulation to represent the forming process of the pottery wheel throwing is developed in using a spring particle system. The hypothetical clay lump is represented with rectangular grid of particles in a particle system wrapped to form a cylinder. This cylinder represents the outer surface of the pulled pot. This cylinder then rotates along a central axis like on a pottery wheel and a digital tool replacing the human hand applies force on to the particles in a variety of ways.
Tool can be defined by 5 parameters
:
1) height of the center of tool from the base of the pot,
2) length of the tool itself,
3) angle it form to the z-axis and on the fixed plane,
4) pressure value, as well as
5) the direction it is being applied.
Each of pulled pot are rotated five times a constant speed. And after every rotation the tool parameters are discretely updated in a randomized manner. As the tool is updated only after each rotation there is radial symmetry maintained as the cylinder gets updated.
The hyperparameters tested to build the ANN model are given below:
a) Size of Data Set - 100, 1000, 10000, 100000 samples
b) Train vs. Validation vs. Test set ratio - 80:05:15, 60:20:20, 70:15:15
c) Batch size - 10, 100, 1000
e) Learning rate in optimization algorithms – 0.001, 0.01, 0.1
f) Choice of optimization algorithm - stochastic gradient descent (SDG), or Adam optimizer
g) Choice of activation function in a neural network layer - Sigmoid, ReLU, SeLU
h) Number of activation units in each layer - -10, 40, 60, 80, 110
Each of the 5 tool parameters are then normalised from 0 to1 to give the 5 (rotation) * 5 (tool parameters) input dimension . The simulation was run 60000 times and using data augmentation 120000 data samples were obtained.
Representing the generated formed pots
The radial distances of the final form after each rotation are normalised from 0 to1 giving 5 (rotation) *13 (form profile) output values. As for the study the interest is in learning the final form, only the 13 output parameters from the last rotation are considered which are highlighted in the table below.
The result above shows that it has considerable influence on the study and ~10000 data samples are required for the neural network models to generalise a function in the parameter space setup in this study.
Input and output for the dataset Input
This gave 25 input parameters representing the tool form 5 successive rotation and 13 output parameters representing the radial distance of equidistant points from the cross-sectional profile of the final pot geometry.
The result above shows the effect on the accuracy of the model due to varying the various hyper-parameters. These parameters were identified which were then used to train the ANN.
Several attempts were made to implement alternative models. These efforts were realized with the aim of improving the efficiency of the model in predicting form based on the tool parameters, while being able to define some relation in the values of successive rotation. Deep-learning models ( an Artificial Neural Network (ANN), a Deep ANN, a Recurrent Neural Network (RNN) and Sequence to Sequence Model (Seq2Seq) model) were used for learning the forming process tried and there layer architecture have been shown below.
Qualitative
Visualizing the predicted form of 9 random samples from ANN training model
Artificial neural network
Recurrent neural network
Deep neural network
Visualizing the predicted form of 9 random samples from DNN training model
Predicting new pots
For predicting new pot, the trained model is presented with unseen input parameters from the test dataset. Its important to note here that the output parameters are interpolated to a curve using the Grasshopper component purely for aesthetic purposes. The Interpolated Curve method takes specific set of points to create the curve, and the resultant curve passes through each of these points, regardless of the curve degree. An example of a sample is shown below in the figure. This step is done for both the predicted parameters and the actual parameters of the form of the pot from the procedural process.
Visualizing the predicted form of 9 random samples from RNN training model
Visualizing the predicted form of 9 random samples from seq2Seq Encoder-decoder training model
As output parameter generates the profile curves of the pots which renders its final form, similarity tests for curves have been used on each of the testing set samples and their distribution is plotted in box plot for comparison. We can see one profile of a single sample taken form the ANN testing dataset shown on the side. And below are various curve similarity tests and their results.
Low losses with low accuracy translate to many small error . Metrics like loss and accuracy in respect to regression tasks where the function approximator looks over continuous values, accuracy as a metric is not a correct way to check for model performance. And thus, the actual values were plotted against the predicted values both shown on side. Hence the values of the performance metrics correlate with the deviations a seen the actual vs predicted plot. The results form ANN and DNN is able to approximate the form but smoothens it out. On the other hand, model based on RNN and the Seq2Seq model does capture the details which is seen more prominent in the later, however not accurately. The details seem to have been wrongly reproduced which in fact visually makes it seem more incorrect than the renders form the first two models, as a bulge where a dent is supposed to be stands out immediately to the human eye. This is again missed in the curve similarity analysis. Hence a better metric to capture the pot profiles is needed.
Similarity measures in models where the blue is for ANN, Orange for DNN, Grey for RNN, and Yellow for Seq2Seq model
The subject of knowledge transfer has been addressed, and it can be argued that experiential knowledge in making can be integrated in digital systems for the purpose of automated fabrication. The methods proposed in this study can be used as an aid and extended upon by training the model with real data from pottery throwing experiments. And such a tool can be used to develop idiosyncratic styles within the digital tool that is limited in the use of CAD/CAM. Such a model can be combined with a setup from other works like that by Charlotte Nordmoen s robotic designer (as shown on the right) to develop new pots capable of replicating craftsmen s styles.
develop multi-material system that rely on pneumatic deformation to carry out shape change
Introduction
type Studio Team Project time 3 months | 2022
keywords material behavior, cyber physical
In the field of adaptive architecture traditionally rigid member deformations are used to achieve state change. However, the project took inspiration from other domains of soft robotics and textile design. Using soft membrane and pneumatic action combined with technique from textile design provided light weight solution for achieving programable structure which occupy less volume when deflated. Adding to it they offer the occupants to feel safer while interacting as compared to what rigid members provide.
To design multi-material system that deform with pneumatic actuation and study various patterns and their effect on resulting deformation.
Behavior and properties of material were considered by making different physical models by keeping the pattern same. Various inflating membranes with bending members on the edge and crease patterns were tested as seen below. Once an appropriate material was chosen, the behavior driven by topology and geometry had to be explored.
To aid the development of an inflatable adaptive material system a cyber physical system is designed which can track and provide feedback to the digital model to improve its performance. As in cyber-physical systems, physical and software components are deeply intertwined and thus require an interface to be able to operate on different spatial and temporal scales based on context.
The overall goal was to deform from 2D patterns into 3D forms, producing distinct deformations and shaping space. So basic forms were studied. In the circles it was aimed to produce double curved deformation; for squares, distinct curved edges; for petal shapes, an enclosure was tried out.
To generate and test out geometric configurations, a digital pipeline to simulate the material system was built which allowed to quickly test out various topologies and their resulting behavior s digitally.
Simulation
Control Crease Type Exploration in Simulation
Basic Form Explorations
Control Pattern Exploration in Simulation
Form exploration
One of the selected pattern was built physically in two different scales to identify the scalability of the material system.
This required comparing the spatio-temporal performance to that of the physical model and the simulation one to improve it such that the digital tool can be used to not only arrive at new forms but also in future to control its distinct behavioral modalities, and interact with in ways that change based on context.
Two different tracking systems were explored which could inform 1) the stable state of the system, and another 2) informing the dynamic shape change with real-time tracking. The devices could work withing the smaller scale of 2m x 2m x 2m object dimension. The feedback loop which is essential for updating the digital model in real-time remains to be developed.
point cloud geometry real-time tracking mesh
physical
developing physics environment for spring system based soft bodies and optimizing for behavior using GA
Challenge
In the field of soft robotics, physical agents are built which conform to their surroundings and mimic biological actuation. One of the key focus of soft robotics is on the kinematics and dynamics through material properties to enable interactions with environments. And its a design problem in itself to find the form which can achieve the desired dynamic behaviour.
This project explores the use of Genetic algorithm to arrive on the configurational geometry given a desired behaviour.
The soft body system is considered comprising of nodes such as air poc kets taken as inspiration from other projects. However most of these projects are either bistable or have limited movement requiring many actuators. Hence a system with 2 actuation is considered. And top down approach is taken to find geometry for driving behaviour.
The soft body is simulated as a spring system and the behaviour (of traversing a plane ground) has been adapted by using Genetic Algorithm.
The spring system comprises of nodes called Joints. Joints are of two type: j1 and j2 which can be thought of two different states of actuation. At t0 jt1 contracts and jt2 expands. And at t1 —j1 expands and j2 contracts. This force along with joint collision avoidance form the internal forces. On the whole spring system there are also externals forces of gravitation and friction which acts to make the body achieve the state at a give time tn.
Then a population of these bodies are created with 10 such individual bodies. Which are then evaluated based on the distance they travel from the origin . The farther they travel, the higher their fitness value. Based on this the best performing creatures are taken and to get the new population. Using a very basic fitness function the desired behaviour of traversing a plane could be achieved .
Structure Generating
t1 t0
joint type 1 joint type 2
point pooling (random walker)
Generations of soft bodied creature
joint + muscle
Genetic Algorithm
larva (Delaunay mesh)
creature (recalculate tri mesh)
internal forces (spring constant, internal collision detection)
Obser vations
genotype
1D array joints with position & type
breeding finding a mid anywhere in the array and crossing over from the the another
All t e bodies mostly move to t e rig t due to ardcoding t e distribution of j1 and j2 to approximately t e alf of t e joints type along t e 1D arra
Currently t e GA evolves into snake like forms to maximize t e lengt and ence extends t e center of mass.
T ere is no scope for particular feature retentions.
T e particle system can be tested against more varied environments like curved sur faces and wit external forces like wind or water.
T e GA can be improved upon by adding constrains on t e fitness function t at rewards more connected masses. And mutation to consider group of joints rat er t an singular joints. Also ot er be aviors like jumping , rolling and forming particular s apes can be ac ieved.
mutation for 5% joints generate a new random position and type
fitness distance from origin to center of mass
Crossover being rudimentary, removes parts t at make t e body perform better itness increases and t en plateaus around a value. T is is currently due to t e mutation on random joints but overall t e form converges and doesn t c ange muc .
Overall t e system can be extended to 3D and t e joints rat er t an being points can be a network of sp ere w ic in application can be replaced by inflatable pockets. And in principal wit a single pump actuation can bring about dynamic be avior s in t e inflatable structure.
mesh generation and panellization based on input vectors
type Academic assignment time 1 months | 2022
keywords parametric, mesh paneling, C#
The roof panels are oriented based on the angle of sun vectors to the ground at a given time. As it changes, so does the panel’s orientation and thus opening.
Based on given site area 1) the script generates a hypar surface which defines the roof structure. And then 2) takes given sun direction (vectors) on the site and reads them one by one and passes it to 3) then generate panels for the roof the number of which can be controlled. The openings of the roof panels at a given time are determined by the sun vector and an attractor (shown in yellow sphere). Finally, 4) the mesh faces are serialized to be exported to a CSV file which can further be imported into a combinatory model.
The script gives controls to the designer to change the curvature of the surface along with the degree of curvature of the surface as shown below. And based on the surface is then the designer can choose the number of sub-division of panels on the surface (as shown on the right).
The attractor shown in yellow controls the depth of the panel opening. Away from the attractor point the panels open more by increasing their depth and nearer to the attractor point the panel stay closed.
mesh generation and panellization based on input vectors
type Workshop - team of 4 time 3 days | 2022
keywords form finding, gridshell, prototype
The workshop introduced the concept of minimal surfaces and involved an exercise in form finding using provided tools. Initially, using Maya low polygon shape for staircase unit was determined. From there the triangulate surface was used as an input within Rhino+Grasshopper. A custom tool then takes the input surface and generates the form found surface based on minimal surface phenomenon where the mean curvature is zero. This surface is then further used to get the asymptotes curves and the grid based on user parameters. The final geometry output was then tested by building a prototype in wood.
Shape Modelling
low polygon modelling
triangulated mesh
Fabrication
https://www.pallaviray.com/