Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Dissertation Proposal
Embedded Actuators, Sensors, and Structure in Adaptive and Distributed Compliant Systems
- An Engineering Analogue to the Musculo-Skeletal-Nervous System -
Brian Patrick Trease
May 23, 2005
The University of Michigan Dept. of Mechanical Engineering
Ann Arbor, MI 48109
Thesis Committee Sridhar Kota, Chair
Brent Gillespie Kazu Saitou
Paul Webb
ABSTRACT Electromechanical systems are generally comprised of separate, nearly rigid components, with relative motion occurring at discrete locations, e.g. joints. In contrast, many designs in nature develop as one connected whole, with flexure distributed throughout the material. In particular, 90% of all living creatures, the invertebrates, rely on compliance for motion and force transmission. Compliant mechanisms are a relatively new class of mechanisms that use engineered elasticity of the constituent elements to transmit motion and/or force. Since flexure is permitted in these mechanisms, they lend themselves to integration with unconventional actuation schemes including thermal, electrostatic, piezo-electric, artificial muscle, and shape-memory-alloy actuators. Here I propose automated synthesis of compliant systems, defined as joint-less monolithic mechanisms embedded with actuators and sensors. I will develop methods for optimal topology, size, and shape of monolithic elastic mechanisms with integrated synthesis of sensors and actuators, for maximum energy efficiency and adaptive performance. The proposed research lays a scientific foundation and a mathematical framework for systematic synthesis of biomimetic designs by addressing issues fundamental to a new paradigm of distributed sensing within a compliant active structure. The design methods and the applications of compliant mechanisms apply to many domains in micro, meso, and macro scales. Engineers and nature both prefer robustness, adaptability, and autonomy in their ‘designs’, yet no paradigm exists for the structured practice of biomimicry. I recognize that the goal of biomimetics is still an open research question to be explored. The motivation for this proposal is to address fundamental research issues in compliant mechatronics towards development of a systematic synthesis methodology and a practical demonstration. Through this proposed research, I will investigate basic research issues including (i) optimal layout of the structure, actuators and sensors and (ii) the groundwork for embedded controls in adaptive compliant systems. I propose a generalized synthesis scheme in which the design requirements are captured in a mathematical form to transform an initial grid of elements into an optimal layout of elastic beams, sensors, and actuators. I will employ genetic algorithms for their ability to find global optima for discrete, nonlinear problems. In addition, the actuators will be mathematically characterized by their operational load-deflection curves and active stiffness, in contrast to the practice of simple point-load representation. The broader impact of the proposed research is: (i) high performance designs of several products ranging from autonomous robots, surgical instruments, adaptive wings, structural health monitoring systems, and adaptive orthoses, (ii) a deeper insight and appreciation of biological systems, and (iii) advancement of the state of the art to the next level of sophistication through creation of monolithic systems with embedded-and-distributed sensing and actuation, in contrast to individually designed structures, actuators, and sensors.
Brian Trease, University of Michigan Page 2 of 63 6/9/2005
TABLE OF CONTENTS
ABSTRACT .................................................................................................................................................................2 1. INTRODUCTION..............................................................................................................................................5
1.1. MOTIVATION...............................................................................................................................................5 1.2. PERSONAL MOTIVATION .............................................................................................................................6 1.3. APPLICATIONS.............................................................................................................................................7 1.4. BENEFITS.....................................................................................................................................................7
2. PROBLEM STATEMENT / HYPOTHESIS...................................................................................................8 2.1. VISION OF AN ADAPTIVE COMPLIANT SYSTEM............................................................................................8
2.1.1. Embedded Actuation ............................................................................................................................10 2.1.2. Embedded Sensing ...............................................................................................................................11
2.2. THEORETICAL FRAMEWORK FOR THE PROPOSED STUDY............................................................................11 3. SCOPE ..............................................................................................................................................................12 4. BACKGROUND / LITERATURE REVIEW ...............................................................................................12
4.1. COMPLIANT MECHANISMS ........................................................................................................................12 4.1.1. Distributed Compliance.......................................................................................................................13 4.1.2. Synthesis Methodology ........................................................................................................................14 4.1.3. Current Research Directions and Labs ...............................................................................................14
4.2. LOAD PATH METHOD ................................................................................................................................15 4.3. RELATED WORK (ACTUATOR ARCHITECTURE).........................................................................................17
5. BASIC RESEARCH ISSUES..........................................................................................................................19 6. RESEARCH PLAN OVERVIEW ..................................................................................................................19
6.1. SPECIFIC RESEARCH ACTIVITIES ...............................................................................................................22 7. PRELIMINARY WORK AND RESULTS....................................................................................................22
7.1. STUDIES WITH GROUND STRUCTURE APPROACH AND GENETIC ALGORITHMS..........................................22 7.2. TEST STUDY PARAMETERS........................................................................................................................26 7.3. TEST STUDY RESULTS ...............................................................................................................................30 7.4. CURRENT CHALLENGES: DENSE STRUCTURES ..........................................................................................33 7.5. MODIFIED GENETIC REPRODUCTION.........................................................................................................35 7.6. THE QUESTION OF CONVERGENCE ............................................................................................................38
8. DETAILED RESEARCH PLAN....................................................................................................................39 8.1. DESIGN ALGORITHMS ...............................................................................................................................39
8.1.1. Task Definition.....................................................................................................................................39 8.1.2. Problem Discretization ........................................................................................................................40 8.1.3. Nonlinear Topology Optimization and Synthesis.................................................................................43 8.1.4. Objective Functions .............................................................................................................................44 8.1.5. Constraints...........................................................................................................................................44 8.1.6. Cross-over strategies ...........................................................................................................................45 8.1.7. Genetic Engineering ............................................................................................................................45 8.1.8. Size / Shape Optimization ....................................................................................................................45 8.1.9. Finite Element Confirmation ...............................................................................................................46 8.1.10. Physical Prototype ..........................................................................................................................46 8.1.11. Integration of Supplementary Concepts..........................................................................................46
8.2. ACTUATION...............................................................................................................................................50 8.2.1. Actuation Selection and Evaluation.....................................................................................................51 8.2.2. Characterization ..................................................................................................................................51
Brian Trease, University of Michigan Page 3 of 63 6/9/2005
8.2.3. Commercial Artificial Muscle Actuators .............................................................................................52 8.3. BIO-DISCUSSION – “BIOMIMETICS”...........................................................................................................52
8.3.1. Observations & Comparisons..............................................................................................................52 8.3.2. Inspiration via Direct Borrowing ........................................................................................................53 8.3.3. Inspiration via Objective / Task...........................................................................................................54 8.3.4. Genetic Algorithms ..............................................................................................................................55 8.3.5. Idea of an integrated Musculo-Skeletal-Ligament-(Nervous) System..................................................55 8.3.6. Biological Evidence of Distributed Actuation .....................................................................................55
9. CONFIRMATION OF RESULTS .................................................................................................................56 10. CASE STUDIES ..........................................................................................................................................56
10.1. VARIABLE STIFFNESS ................................................................................................................................56 10.2. ADAPTIVE ORTHOTIC AND PROSTHETIC DEVICES .....................................................................................58 10.3. SHAPE-CHANGING WINGS AND FINS...........................................................................................................58
11. RESEARCH TIMELINE ...........................................................................................................................58 12. POTENTIAL CONTRIBUTIONS.............................................................................................................59 13. SUMMARY .................................................................................................................................................60 14. REFERENCES............................................................................................................................................61
14.1. GENERAL...................................................................................................................................................61 14.2. COMPLIANT MECHANISMS ........................................................................................................................61 14.3. RELATED WORK........................................................................................................................................62 14.4. LOAD PATH METHOD ................................................................................................................................62 14.5. BIOLOGY ...................................................................................................................................................63
Brian Trease, University of Michigan Page 4 of 63 6/9/2005
1. Introduction
1.1. Motivation The research topic I propose is a new methodology for the synthesis of distibuted
compliant systems with embedded actuation. The basic premise of a compliant system is the
integration of motion/force transmission via elastic deformation with embedded actuation and
sensing. The motion and force transmission problem has been researched for more than a decade
at the University of Michigan. A new field of jointless monolithic devices called Compliant
Mechanisms (CM’s) has been established, with Michigan’s Compliant Systems Design
Laboratory (CSDL) authoring numerous papers and software on the topic. These devices use
flexure and deformation to transmit motion and force, rather than rigid bodies with conventional
mechanical joints. This design paradigm is inspired by nature, where strength and compliance
are observed in structures, as opposed to the goal of strength and stiffness employed in traditional
engineering. The proposed research extends this biological inspiration to create truly monolithic
systems – autonomous, adaptive, efficient, self-contained devices.
I propose to further explore the potential and logical ends of compliant mechanism
technology. There is an interest and a need to synthesize compliant mechanisms with elements
of other fields including controls and adaptive structures. By combining these fields, a new
paradigm emerges for complete and embedded systems that offer a basis for autonomy. In doing
so, we continue to be inspired by nature, and become more intimate with the question of the
relationship between nature and engineering design. If indeed we feel compelled to take
inspiration from biology, then we need to better understand why and how we can appropriately
use biomimetics.
The objective of the proposed research is restated as the further advancement along the
path from traditional mechanical design forward to biomimetic design, as graphically depicted in
Figure 1. The steps made will include systematic design tools and algorithms. From these it is
also desired to gain the specific insights that can lead to general engineering design guidelines
for embedded systems.
Brian Trease, University of Michigan Page 5 of 63 6/9/2005
Figure 1: Transition from Traditional Mechanical Design to Biomimetic Design with Distributed and
Embedded Compliant Systems. Current State-of-the-Art is Compliant Design.
1.2. Personal Motivation The development of a new methodology poses completely new challenges in formulating
the synthesis equations and objective functions to yield compliant systems. While the topic is an
interest of my advisor, my previous research experience both prompted the question of
embedded actuation in my own mind and provided a path and tools for investigation. I have
been studying compliant mechanisms for five years, first considering their implementation as
macro-scale adaptive structures. Several times I have addressed the problem of forming a
mathematical basis for motion synthesis via such mechanisms. I have also explored this at the
micro-scale for internally-actuated electro-thermal MEMS devices. I have worked on CM
synthesis projects using both continuous and discrete optimization strategies. I have also led a
project on the design of a biomimetic aquatic vessel with compliant-fin stroke propulsion,
provoking my thought in the usefulness of biomimicry. Most recently, I am part of an effort to
integrate compliant mechanisms with haptic control devices.
Further, I have several years experience with a variety of smart-material actuators. I have
observed that actuation type, placement, and magnitude are always design parameters defined in
the project statement for CM synthesis. Having now worked on many projects based on the
synthesis of motion from elastic mechanisms, I propose to move the technology one step forward
in my Ph.D. research. By including both actuators and sensors within the synthesis algorithms
for new devices, an already promising technology will be made relevant to a broader range of
applications.
Brian Trease, University of Michigan Page 6 of 63 6/9/2005
1.3. Applications The use of multifunctional elements as both structural members and actuators, the
integration of multiple materials, and the freedom from dependency on external drives comprise
a biomimetic nature specifically attractive to several exciting fields. Many biomedical
applications require these qualities. Examples include surgical tools and grippers, active-assist
joints, and active valves, all at the millimeter scale. Another area where compliant mechanisms
have already proven to perform well is in MEMS. Often, the only practical solutions for MEMS
design problems have been delivered via compliant mechanisms. As research of smart-material
implementation at the MEMS level increases, this proposed research will advance the creation of
autonomous devices at the micron scale, unfettered by external actuation sources.
Assuming perfect success of the methodology, a mathematical framework for design of
smart/adaptive systems will be enabled. The embedding of sensing, actuation, and control will
aide in the design of autonomous vehicles. Smart structure applications include adaptive
orthoses and prosthetics and conformal interfaces.
1.4. Benefits The primary benefits of this research are at the product and design levels. Embedded and
distributed compliant systems will provide many advantages in applications requiring robustness
and autonomy. Energy efficiency can be increased, reducing requirements of external energy
sources. Additionally, internalization of the actuator and sensor components offers protection
from the external environment.
At the design tool level, our research will benefit those seeking methods for tailored
design of synergistic, monolithic systems. Increased actuator efficiency can be attained by
seeking global optima at the system level. These new methods will be systematic and applicable
to a general case of adaptive structure problems.
Secondary benefits of this research occur at the learning level – the study of biomimetics.
Again, one of the benefits of this research is in exploring the potential to learn from nature.
Delimiting the potential role of biomimetics results in understanding how biological insights are
useful to engineering. Indeed, many comparisons can be made, but which of these provides
benefit and useful knowledge? Arguments can be made for direct mimicry or for the indirect
borrowing of apparent principles and objectives in biological design. To add credibility to the
Brian Trease, University of Michigan Page 7 of 63 6/9/2005
idea of biomimicry, we should identify the cases in which nature should and should not serve as
engineering inspiration.
2. Problem Statement / Hypothesis
The proposed activities of this research are part of a much larger vision. The paradigm of
distributed systems poses many new challenges in the function and architecture of structures,
actuator, and sensors. The grand research vision, described in this section, is to investigate and
gain fundamental understanding of the basic research issues underlying those challenges. The
focus is then narrowed in the Scope and Research Plan Overview sections.
I hypothesize that modified versions of the existing algorithms for compliant mechanism
optimization can be implemented to design efficient and controllable systems of integrated
structures, sensors, and actuators (see Figure 3). The physical layout of these components will
be optimized for energy efficiency, controllability, and responsiveness, while satisfying various
weight and performance constraints (e.g. for autonomous robots). Figure 4 shows an example of
the problem definition and a conceptual solution. Allowing for multiple inputs in compliant
mechanism design leads to the question of control. A method for including knowledge of the
controller during the design synthesis may results in more controllable and sensitive structures.
It is also intended to increase the accuracy and functionality of previous compliant
mechanism design schemes by including the actuator and its stiffness in the optimization for the
first time. This provides a more realistic view of actuators, which provide neither a fixed force
nor a fixed displacement, but rather a load-curve relationship for a given power input. Whether
stiffness-matched systems will naturally result from the simultaneous optimization of structure
and actuator is a pertinent research question. Further, I seek to develop a general, parameterized
embedded actuation system methodology, for which any particular actuator's characteristics only
need be specified. The final outcome will be the transformation of computer-generated designs
with optimized geometry, materials, and properties directly into the physical realization of
distributed compliant systems.
2.1. Vision of an Adaptive Compliant System Engineers generally mold electromechanical systems in the rigid-and-discrete paradigm:
a rigid structure with mechanical joints is first designed and then actuators and sensors are
Brian Trease, University of Michigan Page 8 of 63 6/9/2005
integrated with a final afterthought of designing controls. In spite of the “mechatronic”
revolution of the early 90s, this paradigm is still prevalent today. I propose a compliant-and-
distributed paradigm with embedded (not integrated) actuation and an embedded sensing
continuum. In this research, “embedded” refers to systems simultaneously designed for optimal
performance. “Integrated” refers to components that are “slapped on” to an already completed
design. Exploiting compliance or elasticity of the underlying structure has several benefits
including elimination of mechanical joints, joint-wear, and joint-clearance. The proposed
paradigm also embodies the notion of “continuous”. Rather than placing discrete sensors
throughout a rigid and discrete structure, distributed sensing can be achieved by embedding
compliant sensors (simple wires coupled by mutual inductance) within a compliant structure (see
Figure 4). Such an arrangement may be practical for many applications ranging from haptic
feedback on surgical manipulators to sense-and-control adaptive aircraft wings.
The vision of an adaptive, embedded compliant system is a synthesis of compliance,
actuators, sensors, and controls. It is a system that both senses and responds to the external
environment via internal structures; a black box converting inputs to outputs. This trade from
external to internal components offers several benefits. First is efficiency by part reduction. A
few well-placed actuators inside a structure may be capable of many complex deformations of
the external structure. Likewise for sensing: rather than numerous, costly pressure sensors along
the external boundaries of the system, a small number of internal sensors can determine the
internal state of stress and map it to the external environment (see Figure 3 and Figure 4). Such
elements are safe from potentially harsh environments; they are even protected from the
conditions they are sensing or imposing.
Insulator(Electrical
and/or Thermal)
Structuraland
EmbeddedActuators
andSensors
Biocompatible
Ext
erna
l Sen
sing
DesiredOutput
Figure 2. Sample design domain for an Adaptive Compliant Systems Problem
Brian Trease, University of Michigan Page 9 of 63 6/9/2005
Flexible Structure
Ext. SensingActuators
Biomaterial
Insu
late
d Ba
se
Internal Sensing
Figure 3. Artistic Vision of a fully Embedded and Distributed Compliant System
The idea of internal, remote, and local components performing external and global
sensing and actuation is biologically inspired. In human and animal biology/psychology, distal
attribution is the experience of thinking we are actually sensing phenomena directly where they
occur, whereas we usually sense them in structures internal to the body (Loomis, 1992). While
we speak informally of "the five senses" that connect us with the external world, most of our
senses are in fact monitoring internal conditions (email corr. with Peshkin, 2005). When enabled
with biomimetic “internal state sensors,” will machines benefit by being so introspective as well?
external pressure sensing
actuators
internal stress state sensors
Figure 4: Artistic Vision of an Internalized and Adaptive Compliant System
2.1.1. Embedded Actuation By “embedding actuators” within a compliant mechanism, I refer to a topology design
process that simultaneously determines the configuration of the actuators and the compliant
Brian Trease, University of Michigan Page 10 of 63 6/9/2005
structure. That is, certain beams in the initial set will transform into actuators during topology
synthesis. The result is not simply an optimized mechanism, but an optimized system with the
number, size, location, and orientation of all actuators matched to the synthesized structure of the
mechanism. Such a system will have better performance compared to one in which the actuator
is “integrated” after the fact. The proposed work lays a scientific foundation for designing
monolithic mechanical transmissions embedded with unconventional actuators such as artificial
muscle actuators.
2.1.2. Embedded Sensing Aspects of the dynamics that are unmodeled or unsensed often limit the performance of
traditional (lumped-parameter) machines under closed loop control. For instance, in a robot arm,
one may have a rotational encoder at a joint, but no way to monitor the flexing of a link. Robot
links are built to be relatively stiff, so that modeling them as perfectly stiff is a good
approximation, at least up to some frequency. Stiffness adds weight of course.
Compliant machines with distributed compliance, inertia, and sometimes damping or
frictional elements, are more challenging to model or sense than lumped-parameter machines.
Whereas traditional mechanisms use point-like sensors (e.g., the joint rotation sensor in a robot),
compliant mechanisms do not concentrate their motion at a single point (c.f. a revolute joint).
The motion that must be sensed is distributed spatially across the entire extent of a compliant
member. This "distributed" character of motion in a compliant mechanism is well matched to
the "distributed sensing” which is characteristic of mutual inductance sensors. By proper design
of the mutual inductance sensors (email corr. with Peshkin, 2005), we can effectively integrate
the motions of the compliant member being sensed over its spatial extent.
2.2. Theoretical framework for the proposed study While many methods may be suggested to explore the design of adaptive, biomimetic
structures, I propose to primarily work within the framework of structural design optimization. It
is an area well-suited and proven in structural and mechanism problems, as described in the
Background section. It promises to fare well for the inclusion of actuation, sensing, and control
in the design objectives. Several objectives will be formulated to explore the problem. These
fitness functions that drive our problems will be formulated from energy principles, observations
of nature, and aspects of control theory.
Brian Trease, University of Michigan Page 11 of 63 6/9/2005
3. Scope Specifically, the scope is to formulate a topology synthesis optimization algorithm for
compliant structures that includes structures, actuators, and sensors. In this research, I will focus
on the pseudo-static cases. Because dynamic, time-dependent phenomena are very important for
some tasks, for now I take care to correctly lay the groundwork so that such items can be
considered in the future. In setting the framework for sensing and control, I will define for
structures the ideas of observability and controllability. These definitions will be used to
evaluate these quantities within the design algorithm.
Concurrent with the development of a design methodology, several specific areas of
biology will continue to be studied. These include variable stiffness in structures and its various
uses, shape-change for locomotion and/or attenuation, and closed-loop control in force or
displacement feedback systems.
The creation of new optimization algorithms will mark completion of the research. I will
show several examples to evaluate these algorithms, including some variable-stiffness designs
and several case studies for load-adaptive airfoils and fit-adaptive orthoses. Further, while it is
difficult to predict the results of exploratory research, I aim to describe at least one new aspect
about the relationship between nature and engineering.
4. Background / Literature Review
4.1. Compliant Mechanisms Compliant mechanisms are of immediate appeal in addressing the problem of complex,
biomimetic deformation because of their inherent flexibility and distributed compliance. The
synthesis of these mechanisms has been well studied, and their advantages have been well
documented, including: absence of wear, backlash, and friction and ease of manufacture. The
pioneering concept of distributed compliance led to development of design algorithms for the
creation of compliant systems. The design methods and the applications of compliant
mechanisms apply to many domains in micro, meso, and macro scales.
Brian Trease, University of Michigan Page 12 of 63 6/9/2005
(a) (b) (c)
Figure 5: (a) Comparison of a multi-part conventional office-stapler with a no-assembly compliant stapler, (b) Adaptive Compliant Wing – embedded compliant mechanism provides leading edge camber change on demand (c) A MEMS electrostatic actuator integrated with a compliant motion amplifier running at 27KHz.
4.1.1. Distributed Compliance Figure 6 illustrates the idea of distributed compliance, showing a MEMS actuator with
integrated compliant motion amplifier.
Figure 6: A MEMS actuator integrated with a 20X motion amplification compliant mechanism is being mechanically probed. Figures illustrate distributed compliance where beams and the mechanism deform as a whole without any stress-prone flexural joints. The device was fabricated at Sandia National Labs using Sandia’s SUMMiT-5 process.
It is important to note a significant difference between the compliant mechanisms
discussed in this proposal and conventional flexures. Conventional flexures have relatively rigid
sections connected by very thin flexural joints. These flexures localize the deformation and are
prone to high stresses and reduced fatigue life. Such flexures have been known for a long time
(Paros & Weisbord, 1965) and have been successfully employed in less-demanding applications
(ex. shampoo bottle lid). The mechanisms discussed here have distributed compliance. That is,
the mechanism deforms as a whole and does not have any flexural joints. In fact, the links or
beams deform without any connecting joints.
The challenge in designing distributed compliant mechanisms is to determine the optimal
distribution of the material to fit within a given space such that the resulting design not only
performs the desired kinematic function but also satisfies various constraints such as permissible
stresses, fatigue, stiffness (natural frequency), manufacturing (minimum widths and thicknesses),
weight, and power.
Brian Trease, University of Michigan Page 13 of 63 6/9/2005
4.1.2. Synthesis Methodology Synthesis involves two steps: (I) generation of the mechanism topology and (II)
determination of optimum size, geometry, and shape of various constituent elements of the
mechanism. A pictorial summary of one such design method is shown in Figure 7, illustrating
how a compliant gripper is created algorithmically. Starting with functional requirements of
desired forces and displacement, a conceptual design is automatically created in Stage I topology
synthesis. Based on material constraints (i.e. permissible stress, strain), fabrication constraints
(minimum feature size, etc.), external loads, and desired mechanical advantage, the exact size,
shape, and geometry of each of the beam elements is optimized in Stage II.
Figure 7: An illustration of the typical two-stage approach for compliant mechanism synthesis:
(a) Stage I: topology synthesis, and (b) Stage II: dimensional synthesis.
4.1.3. Current Research Directions and Labs A brief history of compliant mechanism design begins with the work of Ananthasuresh in
1994, who first proposed using a structural optimization approach for mechanism design. Many
followed, implementing the ground structure approach, energy efficiency formulation, refined
size/shape optimization, and inclusion of nonlinear deformation effects. Refer to Lu’s
dissertation (2004) for a detailed history.
Brian Trease, University of Michigan Page 14 of 63 6/9/2005
Current research in the field has spread to a small number of universities. At The
University of Michigan we continue to focus on medical applications, biomimetics, new
synthesis formulations, and easy-to-use design software. Research at Brigham Young University
under Larry Howell centers on the Pseudo-Rigid-Body-Approach to compliant mechanisms,
which uses empirical data to employ a traditional kinematics-based approach to mechanism
design. (http://research.et.byu.edu/llhwww/) Mary Frecker at Penn State continues research into
many areas including medical devices and reconfigurable structures.
(http://www.me.psu.edu/ODACSL/) Current research at the University of Illinois in Chicago is
focused on the development of novel MEMS-based miniaturized devices, led by Laxman
Saggere. (http://www.uic.edu/labs/microsystems/index.html) Kerr-Jia Lu is continuing her
research into Shape-Morphing Compliant Mechanism at Geroge Washington University.
(http://www.seas.gwu.edu/~kjlu/) Just Herder at the Delft University of Technology works in
the area of statically-balanced compliant mechanisms, with a focus on medical tools and human-
assist devices. (http://mms.tudelft.nl/staff/herder/index.htm) Finally, Martin Culpepper heads
the Precision Compliant Systems Lab at MIT, where research focuses on MEMS devices,
machine tools, layered and formed actuator technology, and design tools for capture of designer
intent. (http://psdam.mit.edu/) The potential contributions that the research described in this
proposal will make to the literature are listed at the end of this proposal.
4.2. Load Path Method The Load Path Method was developed by Lu to address the issue of structural
connectivity in compliant mechanism synthesis problems. In her dissertation, she observes:
“Since the binary topology variables are defined in the element level, they do not explicitly contain any information about the overall connectivity, which is in the structural level. When certain elements are eliminated at the same time, the GA can produce invalid designs such as those shown in Figure 8 that include disconnected substructures or are disconnected from the input or ground. In general, the structural connectivity is unknown when simply looking at the design variables.”
Brian Trease, University of Michigan Page 15 of 63 6/9/2005
Figure 8: (a) The binary ground structure used to discretize the design domain; (b) an invalid
design with disconnected substructure; (c) an invalid design that is disconnected from the ground; (d) an invalid design where the input is disconnected. (Lu, 2004)
To resolve this problem, a graph-based discretization is formulated as an alternative to
the Grounded Structure Approach. This method may be very relevant to this research because of
the discrete nature of the actuator variables. The Load Path Approach begins with numerous
subgraphs (paths) that already connect all of the relevant points of interest in the design space
(see Figure 9). Thus, the resulting structure is guaranteed to be properly connected. In addition,
the method treats the sub-graphs as variables (not the individual elements), leading to a large
reduction in the total number of design variables and reducing computation time. This process is
further described in the Problem Discretization section.
(a) (b)Figure 9: (a) A fully connected graph of a shape-morphing compliant mechanism.
(b) Different port locations render different geometries in the compliant mechanisms.
Brian Trease, University of Michigan Page 16 of 63 6/9/2005
4.3. Related Work (Actuator Architecture) There are a few literature references to the optimization of actuator architecture, which is
similar in appearance to the proposed research. However, these methods are of different scope
and employ different methods. Actuator architecture problems aim to lay out the material
makeup of an actuator, but do not address how that actuator relates to the rest of a physical
system. Most of these studies focus on the distribution of only the actuator material, without any
supporting structural elements. First, consider the problem shown in Figure 10, which seeks to
find the optimal layout of actuator material in a flexible airfoil (Anusonti-Inthra and Frecker,
2003). The only non-actuator elements are the skin elements, which are not included as design
variables. The result (Figure 11) shows a network of only actuators. Such a structure may be
suitable when considering the entire device as a single bender actuator, but is highly impractical
in describing implementation of actuators within a mechanical system. E. Silva seeks a similar
goal of distributing piezo-electric actuator material in his dissertation work (1998).
Figure 10: Design Domain Used for Actuation Distribution in a Flexible Airfoil (Anusonti-Inthra, 2003)
Figure 11: Actuator Distribution Optimization Results with a Volume Constraint of 50%
Only one relevant paper has been discovered that takes on the task of embedded actuator
with structure during synthesis. Bharti and Frecker (2003) provide a two-stage method for
topology synthesis of piezo-electric actuators. Their work differs from others in that the
discretized elements can become either “active” (actuator) or “passive” (structural) elements. In
the first design process, they treat all elements as actuators and proceed to optimize the layout.
Brian Trease, University of Michigan Page 17 of 63 6/9/2005
Actuators that are understressed and appear not to be contributing to the objective are converted
to structural elements and the structure is optimized a second time. In an example, they redesign
the flextensional actuator shown in Figure 12. But rather than specify the red area as the
actuator, their method distributes the actuator and structure as depicted in Figure 13.
Figure 12: PZT Flextensional Actuator Problem, shown with internal, fixed actuator
Figure 13: Topology Optimization with all elements shown in different shades according to their relative
importance in providing a maximized output deflection. (Bharti and Frecker, 2003)
The results in Figure 13 are undesirable for the same reasons as those in Figure 11. This
is only a “material distribution” problem and not a mean for integrating actuators within a
structure. Bharti does conjecture further, however, that the structure in Figure 13 could be
interpreted to a form like that shown Figure 14. While this interpretation comes closer to our
concept of actuators distributed within a compliant system, no method or rationality is provided
for how such an interpretation is to be made. The final structure shown is in fact only a concept
for what an interpretation might look like; a quick look at it indicates that the actuator locations
are nonsensical and that the structure would not have the same functionality as that shown in
Figure 13.
Brian Trease, University of Michigan Page 18 of 63 6/9/2005
Figure 14: Possible Interpretation of Figure 13. (Bharti and Frecker, 2003)
5. Basic Research Issues The basic research issues are in identifying and implementing the appropriate objective
functions for synthesis of embedded compliant systems. The choice of objective function may
depend on whether actuator power, actuator number, and/or actuator effectiveness is to be
minimized. In addition, potential actuators for use in such systems should be researched. A
practical choice must be made while considering how an actuator’s limitations constrain the
design goals. Once chosen, the next task is to mathematically characterize actuators for
inclusion in the optimization. Successful results will require guaranteed structural connectivity,
an important issue that will be ensured using graph-based methods.
Another issue is setting the framework for “design for control” of embedded compliant
systems. This will be established by developing specific definitions for the input and output
functions of the structure, followed by their mathematical implementation in the algorithms.
Finally, in terms of biomimetics, the issues are identifying what exactly from nature is
worthy of use as an objective function. Another matter is interpretation of synthesized designs in
terms of what is seen in nature, in search of useful correlations or potential opportunities for
continued research.
6. Research Plan Overview The proposed research consists of three concurrent areas of focus. Primary attention is
given to the development of the optimization algorithms. At the same time, I will continue
research into appropriate actuators for embedded compliant systems and their mathematical
characterization. Also ongoing will be the question of biomimicry. What additional objectives
can we find from nature, and how can we gain additional insight from our results via biological
comparison?
Brian Trease, University of Michigan Page 19 of 63 6/9/2005
Figure 15 provides a pictorial overview of the proposed method of synthesis. It is from
this description that I identify the specific research activities to be undertaken. We begin with
the given problem specifications and design space, depicted in (a). The workspace constrains the
maximum allowed size of the system. Portions of the workspace deemed to be rigid (connected
to either ground or another structure) are indicated. The specific output points (blue circle) and
desired output direction (blue arrow) are shown. Many synthesis problems also include reaction
forces at the output point; either a reaction load or an external spring. Here, the external load is
applied along a loading region depicted in pink. The external load in the case is an uneven
pressure along a line, shown by the black arrows. Note that no input ports are specified; it is
only required that the actuators fall within the physical workspace.
Step (b) shows the workspace discretized using the Ground Structure Approach.
Resolution depends on the number of nodes used and the adjacency of connected nodes. In this
case, no elements span more than two square units. Each element is represented by a discrete
variable that will determine its presence in the optimal structure. The structural and actuator
materials are specified, allowing each element to contribute to the overall structural stiffness.
Step (c) is the result of the topology synthesis optimization achieved with a genetic
algorithm. The result reflects the specified objective function subject to any constraints.
Actuators are indicated in red and elements with embedded sensing in blue. Here, the objective
function is the energy efficiency of the structure, calculated as the work at the output nodes
divided by the sum of the work at all the input (actuator) nodes. The work terms are found by
analyzing the structure with the finite element method. Actuators are represented by more than
point force or displacements; they are designated as structural elements with load-deflection
output curves. The structure is required to meet a minimum displacement at the output node
while bearing the depicted external pressure load.
After the topology is determined and fixed in step (c) the individual elements are resized
and reshaped in step (d), the Size and Shape Optimization. All of the external loading conditions
are still applied, but the objective function may be changed to fine tune the structure’s function,
such as optimal geometric advantage. Stress and buckling constraints are also included at this
stage to check and prevent element failure. Note how the elements adjust in length and
thickness, yet the original topology is maintained.
Brian Trease, University of Michigan Page 20 of 63 6/9/2005
Step (e) shows the synthesized structure in its deformed state. The figure shows two
kinds of sensing: external and internal. The distributed internal sensors are preferred to the
external sensors because they can determine external loads with fewer elements, while not
exposed to the environment. It is intended that the internal and embedded nature of the sensors
provides for simplified control of the embedded actuators. Finally, step (f) shows a physical
interpretation of the synthesized structure. The figure conveys the vision of a monolithic
skeleton with embedded sensors and actuators acting as a cohesive whole.
(a) (b) (c)
(d) (e) (f)
Figure 15: Various steps in the proposed synthesis scheme. (a) design specifications (b) initial array of beam elements as a ground structure (c) optimized topology of beams, actuators, and sensors (d) size optimization (e) deformed position (f) physical interpretation.
Brian Trease, University of Michigan Page 21 of 63 6/9/2005
6.1. Specific Research Activities My specific research intentions are summarized in the following list and further
elaborated in the Detailed Research Plan section.
• I intend to include the orientation, location, and number of actuators as variables in the topology synthesis problem for designing compliant mechanisms.
• I intend to use genetic algorithms as the means of optimization, and will pursue two methods of problem discretization: the Grounded Structure Approach and the Load Path Approach.
• I intend to include the structural and operational loading-curve properties of the actuators in the mechanical analysis used to calculate the fitness function.
• I intend to lay the groundwork for optimal sensor layout and embedded controls in adaptive compliant systems. This requires mechanical definitions for “observability” and “controllability”, which will serve as metrics in the design for optimal control.
• I intend to explore and articulate aspects of the relationship between engineering and nature for potential exploitation in design. This relationship includes comparative design, discovery and sharing of objective functions for optimization, and the particular study of variable stiffness.
7. Preliminary Work and Results
7.1. Studies with Ground Structure Approach and Genetic Algorithms To demonstrate the feasibility of this research, I have created some optimization code and
completed a number of initial studies. Thus far, I have used the Grounded Structure Approach
(GSA) with a genetic algorithm, although a number of connectivity problems can occur when
combining the GSA with genetic algorithms. I have added a number of graph-checking
constraints to resolve the issue, but it is still believed that the Load Path method will further
alleviate these problems. The current implementation of the GSA thus serves as a benchmark to
rate the quality of the Load Path method.
In all of the studies performed thus far, the objective function is to maximize the energy,
expressed in term of a spring-based efficiency, η. That is, while the actuators act against the
structure at the input ports, springs are placed at the output ports to simulate the reaction loads
(See Figure 16).
Brian Trease, University of Michigan Page 22 of 63 6/9/2005
koutdout Fin1, din1
actuator Fin2, din2
ground Figure 16: Basis for calculation of Energy Efficiency
This enables the easy calculation of work output divided by work input in Eq. (1), where
all forces and displacements are measured parallel to the intended direction of motion. Thus, this
formulation does not include output work that does not contribute to the mechanism’s function.
Hetrick (1999) provides a derivation for this formulation, shown in Figure 17.
%100
η
2
≤===∑∑ inin
outout
in
out
in
out
dFdk
WW
WW
(1)
Figure 17: The Input and Output Work History (Spring Formulation) (Hetrick, 1999)
In addition to the energy efficiency, two constraint penalties are added to the objective
function. The first is a constraint that requires the output deflection in the desired direction to be
greater than a minimum value, d0. The second requires the output deflection perpendicular to the
desired direction be less than a maximum value, dtan. The final form of the objective function is
Brian Trease, University of Michigan Page 23 of 63 6/9/2005
shown in Eq. (2), where η is from Eq. (1) and w1 and w2 are relative weighting constants. From
the formulation, the penalties are applied when dout < d0 or dtan > . ⊥outd
[ ])d - (d w )d - (d w η maximize outtan20out1⊥×+×+ (2)
In the code, the workspace is broken down into a grid of nodes which are interconnected
by elements (Figure 18). A design variable exists for every element shorter than a specified
length, and determines whether or not the element exists. In practice, each variable is allowed
one of four discrete values (0, t1, t2, t3). A value of 0 deactivates the element, removing it from
the structure. A value of ti indicates one of three thickness values for the thickness of the beam,
e.g. 1mm, 2mm, or 3mm. In addition, there is a variable for every actuator to be used in the
structure. This variable has a value between 1 and the total number of elements. Its value marks
the element selected to be the actuator in a given structure. This approach guarantees a specified
number of actuators for each problem and will prevent results like those seen in related literature
(See the Related Work section).
Amplifier Problem, 6x6 grid, 250 elements Gripper Problem, 5x5 grid, 168 elements
dout kout
dtan
dout
koutline of symmetry
Figure 18: Design Domains and Discretization for the Test Problems
Brian Trease, University of Michigan Page 24 of 63 6/9/2005
Graph searches are implemented to ensure connectivity of the actuators, output port, and
at least one ground point all within the same graph. Disconnected structures that do not meet
minimal requirements for analysis, such at that pictured in Figure 19, will be rejected
(infanticide) and replaced with a new member. A brief overview of this process is shown in
Figure 20 and described in further detail in the Modified Genetic Reproduction section.
0 50 100
0
20
40
60
80
100
Figure 19: Example of a Disconnected Structure that would be rejected from
the population before the assessment of the fitness function. (white circles are ground, green arrow is output node, red lines are actuators)
Also of major importance has been the implementation of genetic modification. This is
referred to as Lamarckian evolution or genetic engineering. Before the structures are even
analyzed, all excess subgraphs are removed (genetic engineering). This often results in
unanalyzable structures becoming analyzable. Further, elements that are discovered not to be
contributing to the objective function after analysis are removed and not passed on
(Lamarckian).
Brian Trease, University of Michigan Page 25 of 63 6/9/2005
a 0 50 100
0
20
40
60
80
100
b 0 50 100
0
20
40
60
80
100
c 0 50 100
0
20
40
60
80
100
d 0 50 100
0
20
40
60
80
100
Figure 20: Brief Overview of Reproduction. (a) initial random design;
(b) trimming of disconnected elements (genetic engineering); (c) analysis for fitness function; (d) elimination of non-participating elements (Lamarck)
7.2. Test Study Parameters The operational parameters of the genetic algorithm have depended on the size of the
chosen grounded structure. For a workspace grid of 5x5 nodes, there are 168 elements; for a 6x6
grid, there are 250 elements; every element is a design variable. As a general rule, the size of the
population should be greater than the number of design variables. Populations of sizes from 150
to 250 members have been tried for both grid sizes. In all cases, the algorithms have been
allowed to run for 1000 generations. The zero-bias for initial populations has been set to values
from 1/30 to 1/60 (i.e. a 1/N chance of existing). There are three types of mutation that occur,
and their likelihoods are as follows: normal element mutation (16%), mutation toward
elimination (mutilation) (30-40%), and mutation of actuator element choice (15-25%).
Two problems have been studied with the algorithms: a displacement amplifier and a
gripper mechanism. These problems are chosen as they are considered benchmark problems for
compliant mechanism design (refer to Lu, 2004). This proposal describes three results for each
Brian Trease, University of Michigan Page 26 of 63 6/9/2005
problem type (A1, A2, A3, G1, G2, G3). The specifications for each of these are shown in
Figure 18. Lu’s results for these problems with a single-output and a single-input are shown in
Figure 21.
Figure 21: Design Domains and Results from Lu (2004) for Benchmark Problems top = amplifier problem, bottom = gripper problem
Many different amplifier designs have been created in previous research (Joo, Kota, and
Kikuchi, 2001; Kota, Rodgers, and Hetrick, 2001 (patent); Hetrick, Joel and Kota, 2003
(patent)). This compliant mechanism serves as a transmission so that the input displacement is
amplified at the output port. The overall design domain is 240mm by 100mm (9.45inch by
3.94inch). To ensure linear motion at the output port, the design is chosen to be symmetric about
the y-axis. Thus, only the right half of the design domain is considered in the synthesis process.
Brian Trease, University of Michigan Page 27 of 63 6/9/2005
Table 1: Design summary of the benchmark studies by Lu (2004) Amplifier Gripper External Load: 1N 8N Input Displacement: 1mm 2mm Geometric Advantage: 27.6 8.35 CPU time: 194sec (3.23min) 51.48sec (0.86min) Maximum stress: 30.17MPa (4.38kpsi) 34.45MPa (5kpsi) Required input force: -98.7N (-22.19lbf) -85.18N (-19.15lbf)
Estimation of design inputs for the new research problems Input forces chosen for new problems: 90N 80N
Output stiffness chose for new problems: (Fout/dout)
1/~20 = 0.05 N/mm 8/(2*~8) = 0.5 N/mm
The compliant gripper design is also one of the most commonly seen benchmark
problems and has been investigated in many previous literatures (Frecker et al., 1997; Hetrick, J.
and Kota, 2000; Joo, Kota, and Kikuchi, 2000; Saxena, A. and Ananthasuresh, 2001). The goal
is to design a compliant mechanism that deforms to grip an object and has appropriate stiffness
to withstand the reaction force upon gripping the object. The overall design domain is 100mm
by 80mm (3.94inch by 3.15inch). The gripper is designed to be symmetric about the x-axis, so
that the output ends will close by input actuation and grip the hypothetical object placed at
(100,0). Due to symmetry, only the upper half of the design domain is modeled with a single
output point located at (100,20).
Table 1 summarizes the results Lu obtained using the Load Path method and a genetic
algorithm. While she used displacement as an input, the new formulation requires force as input,
which is chosen by picking a value near Lu’s calculated input forces. Similarly, output
stiffnesses for my new problems is estimated from Lu’s output loads and output displacements.
These calculated values are listed at the bottom of Table 1. Table 2 lists the remaining
parameters for the preliminary problems studied for my research.
Brian Trease, University of Michigan Page 28 of 63 6/9/2005
Table 2: Parameters for Test Studies depicted in Figure 22 through Figure 27 Amplifier Gripper A1 A2 A3 G1 G2 G3 units Grid Size 5x5 5x5 6x6 6x6 6x6 6x6 nodes Symmetry Y Y N Y Y Y # of Actuators 1 1 2 2 1 2 Initial Zero Bias 1/55 1/55 1/51 1/50 1/50 1/55 Input Force 90 90 90 80 80 80 N Element Stiffness 2480 2480 2480 2480 2480 2480 MPa Actuator Stiffness 2000 2000 2000 2000 2000 2000 MPa Output Stiffness 0.05 0.05 0.05 0.5 0.5 0.5 N/mm Population Size 130 150 230 200 200 230 Generations 200 1000 1000 1000 1000 1000 Min Output d 20 20 20 16 16 16 mm
Penalty Weight 0.1 0.1 0.1 0.2 0.2 0.1 Max Tangential d symm. symm. 5 4 4 4 mm
Penalty Weight symm. symm. 0.6 0.3 0.3 0.3 Element Mutation 0.16 0.16 0.16 0.16 0.16 0.16 Zero Mutilation 0.3 0.33 0.4 0.4 0.4 0.4 Actuator Mutation 0.2 0.2 0.2 0.23 0.23 0.2 Optimization Results Energy Efficiency 21.4 51.4 12.3 23.3 33.2 25.8 % Output Disp. 21.5 29.1 23.4 5.5 11.1 11.1 mm
The test studies included in this paper and summarized in Table 2 are only a subset of all
the studies completed. Table 3 lists the complete set of studies. The number of test studies for
each case is indicated in parentheses. The key parameters that differentiate these studies are:
• Symmetric / Asymmetric • One / Two Actuators • Remote / Free Actuator Location • Push / Pull Actuator • GA Parameters (Population Size, Generations, Mutation Rates) • Grid Size
Brian Trease, University of Michigan Page 29 of 63 6/9/2005
Table 3: Full Set of Test Studies performed in Preliminary Research
1. Amplifier Problem a. Asymmetric Amplifier
i. Single Actuator (6 push, 1 pull) ii. Two Actuator (3)
b. Symmetric Amplifier i. Single Actuator (2)
2. Gripper Problem a. Symmetric Gripper
i. Full Actuator Domain (4) ii. Remote Actuator
1. Single Actuator (2) 2. Two Actuator (1)
7.3. Test Study Results The figures in this section show the optimization results for each of the test studies
alongside their convergence histories. Each figure shows the complete symmetric structure, but
with the deformed shape shown in only one half. Study A1 shows the first attempt at a
displacement multiplier; the optimization was terminated prematurely at only 200 generations.
Study A2 is nearly the same, but with a 15% larger population size and a full 1,000 generations
of design propagation.
-120 -80 -40 0 40 80 1200
50
100
0 20 40 60 80 100 120 1400
10
20
30Objective Function
0 20 40 60 80 100 120 1400
10
20
30Efficiency, in Percent
→200
Figure 22: Study A1 - Single Actuator Symmetric Amplifier. Efficiency = 21.4%, dout = 21.5mm
The result for Study A2 is topologically identical to that from Lu in Figure 21 (top). The
only difference is the layout of the actuator, which was not optimized in Lu’s study. The two
symmetric actuators in Figure 23 act to create a vertical force as specified for the benchmark
case.
Brian Trease, University of Michigan Page 30 of 63 6/9/2005
-120 -100 -80 -60 -40 -20 0 20 40 60 80 100 1200
20
40
60
80
100
120
0 50 100 150 200 250 300 3500
20
40
60Objective Function
0 50 100 150 200 250 300 3500
20
40
60Efficiency, in Percent
→1000
Figure 23: Study A2 - Single Actuator Symmetric Amplifier. Efficiency = 51.4%, dout = 29.1mm
Study A3 departed from the symmetric design space to explore the synthesis of an
amplifier without a guaranteed vertical output displacement. A penalty is imposed when the
output displacement is more than 5mm in the horizontal direction. The jagged history of the
efficiency indicates the conflict between energy efficiency and this penalty constraint.
0 20 40 60 80 100 1200
20
40
60
80
100
120
0 200 400 600 800 1000-10
0
10
20Objective Function
0 200 400 600 800 10000
5
10
15Efficiency, in Percent
Figure 24 : Study A3 - Double Actuator Asymmetric Amplifier. Efficiency = 12.3%, dout = 23.4mm
The next three cases show the results for the compliant gripper problem. The results for
all of these are topologically the same. An underlying four-bar mechanism can be identified in
each figure, with a rigid extension of the coupler link connecting to the output point (see bottom
half of Figure 25). In all of the cases, the actuator pushes on one of the corners of the coupler
link. In Study G3, the coupler is stiffened with some truss elements.
Brian Trease, University of Michigan Page 31 of 63 6/9/2005
0 20 40 60 80 100-30
-20
-10
0
10
20
30
40
0 100 200 300 400 500 600 7000
10
20
30Objective Function
0 100 200 300 400 500 600 7000
10
20
30Efficiency, in Percent
→1000
Figure 25: Study G1 - Double Actuator Gripper. Efficiency = 23.3%, dout = 5.5mm Underlying four-bar mechanism shown in black.
Note that actuator location was restricted in all the gripper studies to ensure that the
actuators are remote from the output point. Only elements on the left-hand half of the design
space could serve as actuators. The unintended result of this restriction is that most of the
element bending occurs on the left side of the structures. The beams connecting to the output
point are mostly stiff and unbending. This is contradictive to our goal of fully distributed
compliance and signals a need to re-evaluate the implementation of the fitness function.
0 20 40 60 80 100-40
-30
-20
-10
0
10
20
30
40
0 200 400 600 800 10000
10
20
30Objective Function
0 200 400 600 800 10000
20
40Efficiency, in Percent
Figure 26: Study G2 - Single Actuator Gripper. Efficiency = 33.2%, dout = 11.1mm
Brian Trease, University of Michigan Page 32 of 63 6/9/2005
0 20 40 60 80 100-40
-30
-20
-10
0
10
20
30
40
0 200 400 600 800 1000-20
0
20
40Objective Function
0 200 400 600 800 10000
10
20
30Efficiency, in Percent
Figure 27: Study G3 - Double Actuator Gripper. Efficiency = 25.8%, dout = 11.1mm Note the formation of trusses to stiffen the output link.
7.4. Current Challenges: Dense Structures A major challenge thus far has been controlling the results of initial random population
generation and of genetic reproduction. Given purely random chance, most initial structures will
be highly dense and highly connected, with almost every node connected to the structure,
resulting in a nearly rigid-body structure, as depicted in the first frame of Figure 28. To control
this problem in the initial population, the likelihood of an element existing is given an uneven
bias toward zero. The greater the bias, the sparser, and usually the more desirable, are the initial
structures (Figure 28). Zero Bias is represented as a fraction, 1/N (used in Table 2), or as a
percent, 100/N% (used in Figure 28). A value of 1/5 (20%), means that any given element only
has a 20% chance of being used, meaning a structural density of approximately 20%. This
comes at a cost, however: many of the sparse initial structures are not valid because they do not
have enough elements to provide connectivity. This conflict is handled by simply rejecting these
initial members and trying again with another random structure (infanticide). The best initial
structures tend to result when choosing a bias that rejects about 95% of the initial attempts. For
example, when attempting to populate the first generation with 200 members, between 4,000 and
7,000 members are checked before 200 good ones are attained. While this may seem
overabundant and computationally costly, it far outweighs the inefficiency of running an
algorithm where 90% of the members cannot even participate in the objective function.
Brian Trease, University of Michigan Page 33 of 63 6/9/2005
20%
0 50 100
0
20
40
60
80
100
10%
0 50 100
0
20
40
60
80
100
5%
0 50 100
0
20
40
60
80
100
3.3%
0 50 100
0
20
40
60
80
100
2.5%
0 50 100
0
20
40
60
80
100
2%
0 50 100
0
20
40
60
80
100
Figure 28: Effect of Zero-Bias on Structural Density of Initial Population Members In each structure, every element has an X% chance of being included.
The problem appears to worsen during reproduction. While two given structures may be
sparse as they stand alone, their cross-over and mutation more often results in an addition of the
structures, akin to 1 + 1 = 100 (See Figure 29(e,f)). One reason for this problem is that there is
no easy way to relate the individual genetic variables to the structures they create. In biologists’
terms, the genotype does not map to the phenotype in any meaningful way.
The density problem is handled by random “hole-punching” (mutilation) in the offspring
before they are ever assessed for connectivity and fitness. This is achieved by randomly turning
off elements within the child structure. This works well because a few breaks in the structure
can eliminate large sections during the subsequent graph connectivity search and trimming. In
practice, this has been accomplished by giving all the elements in a structure a 20% to 40%
chance of being mutilated. In the future, this problem can be handled by better genetic
engineering of the parents before cross-over. Currently, only those elements in parent members
that contribute nothing to the fitness are eliminated. Often, however, many elements provide
next to no effect on the fitness. That is, if these extra elements are removed, the objective
Brian Trease, University of Michigan Page 34 of 63 6/9/2005
function remains nearly the same. By setting a tolerance for how much “value” each element
has in the fitness function, a filter can be created to eliminate those of little value (Lamarckian).
In a sense, the structure that is left behind is truly sought in the first place, for it provides the
basic functional topology.
7.5. Modified Genetic Reproduction Figure 29, Figure 30, and Figure 31 show two examples of the modifications I have made
to the genetic reproduction process. Figure 29 and Figure 30 each follow two parent structures
through cross-over and mutation. First, the original parent members are shown (a,b), followed
by the children that result from cross-over (c,d). Next is shown the result of mutation and
mutilation (e,f). Finally, the graph checks are performed. An invalid structure (g) is rejected
(infanticide) and a valid structure (h) is trimmed to its essential elements (genetic engineering).
Figure 31 demonstrates the fitness function and post-processing of the surviving structure from
Figure 30. The complete process for genetic reproduction is listed here:
1. Cross-over & Mutation
• Standard method of reproduction in GA.
2. Mutilation
• Random removal of elements in a structure to reduce density.
3. Infanticide
• Rejection of any structures incapable of even performing the fitness function.
• Rejected members are replaced with new members by starting over again at step one until an acceptable member is found.
4. Genetic Engineering
• Trimming of structures to remove elements that will prevent or not contribute to the analysis.
• Helps keep structure clean and sparse for improved reproduction of next generation.
5. Fitness and Survival
• Standard method of analysis and selection in GA.
6. Lamarckian Engineering
• Evaluation of structural members to see which are not contributing to the objective function.
7. Repeat Process for Next Generation
Brian Trease, University of Michigan Page 35 of 63 6/9/2005
PARENT 1
a0 50 100
0
20
40
60
80
100
PARENT 2
b0 50 100
0
20
40
60
80
100
CHILD 1
c 0 50 100
0
20
40
60
80
100
CHILD 2
d 0 50 100
0
20
40
60
80
100
MUTATION & MUTILATION
e0 50 100
0
20
40
60
80
100
MUTATION & MUTILATION
f 0 50 100
0
20
40
60
80
100
(g) THROW AWAY INVALID STRUCTURE
Genetic Engineering
h 0 50 100
0
20
40
60
80
100
Figure 29. First Demonstration of Reproduction Process
Brian Trease, University of Michigan Page 36 of 63 6/9/2005
PARENT 1
a0 50 100
0
20
40
60
80
100
PARENT 2
b0 50 100
0
20
40
60
80
100
CHILD 1
c 0 50 100
0
20
40
60
80
100
CHILD 2
d 0 50 100
0
20
40
60
80
100
MUTATION & MUTILATION
e0 50 100
0
20
40
60
80
100
MUTATION & MUTILATION
f 0 50 100
0
20
40
60
80
100
(g) THROW AWAY INVALID STRUCTURE
Genetic Engineering
h0 50 100
0
20
40
60
80
100
Figure 30: Second Demonstration of the Reproduction Process
Brian Trease, University of Michigan Page 37 of 63 6/9/2005
0 50 100
0
20
40
60
80
100
FITNESS FUNCTION
→
0 50 100
0
20
40
60
80
100
LAMARCK FILTERING Figure 31: Final Processing of Surviving Design from Figure 30(h).
Non-contributing structural members are deleted.
Ideally, I seek to simplify the reproduction process by improving the genetic engineering
performed by fitness assessment. I hope to analyze the effects of the Lamarckian filtering so that
we can tell what elements are likely to be trivial, and then create a general filtering rule so that
we can delete them during genetic engineering, before the fitness function. In addition, the latest
results have provided some explanation for high-density structures resulting from reproduction.
An initial theory was that the cross-over operation was combining what had been insignificant
elements into significant structures, resulting in a dense overall structure. However, it now
appears that the dense structures result from the basic element mutation. At a high mutation rate,
many elements are added to the structure that were previously inactivated, thus cluttering the
design. By limiting mutation to below 5%, it is hoped to dramatically reduce the need for
mutilation and infanticide. An improved process for genetic reproduction would only use the
following four steps:
1. Cross-over & Mutation
2. Genetic Engineering
3. Fitness and Survival
4. Repeat Process for Next Generation
7.6. The Question of Convergence Finally, there is a question of the nature of convergence of the optimization employed
thus far. It is not clearly evident that the GA’s are actually working by passing down useful
genetic information. The theologian’s argument that our evolutionary algorithms are relying
Brian Trease, University of Michigan Page 38 of 63 6/9/2005
only on random chance may be the current case. 250 population members multiplied by 1,000
generations offer 250,000 possible chances for a good structure, especially when combined with
the genetic engineering that acts to improve these structures “with a divine hand.” Further
observance of the inner workings of the algorithms, and not just the results, is required to
evaluate the efficiency of the proposed synthesis strategy.
8. Detailed Research Plan
8.1. Design Algorithms
8.1.1. Task Definition The first step is development of the optimization algorithms. Hypothetical actuators will
be chosen as the inputs to one of the following problems: Single-Input/Single-Output (SISO),
Multiple-Input/Single-Output (MISO), or Multiple-Input/Multiple-Output (MIMO) mechanism
design. These problems are traditional categories of mechanism design that can be applied to
many situations. Work has already been done on the single-output problem with both single-
actuator and multiple-actuator inputs (SISO and MISO). The next task would be the MIMO
problem, which often takes the form of the shape-change problem. Beyond those tasks, I also
propose to study the multiple-load-case problem, which is the generalization of the adaptive-
output problem. That is, with more than one actuator, more than one output can be achieved, and
it is desirable to achieve as many unique outputs as possible with a small number of actuators.
The development of the adaptive-output problem also leads to the variable stiffness problem.
When the functionality of the design algorithms is established, other objectives observed from
nature may be applied.
Once the type of problem is chosen, all relevant quantities are parameterized for use in
mechanical analysis and optimization. An output node is selected, for which a desired
displacement or output force is prescribed. Multiple output nodes are chosen for shape control
applications. Whether single or multi-node output, we also specify an external-load sensing
region. The structure ‘communicates’ with its environment through this interface. The system
must fit within a specified 2-D design space.
Brian Trease, University of Michigan Page 39 of 63 6/9/2005
8.1.2. Problem Discretization With the actuator chosen and the problem defined, the next step is to discretize the
problem so that it may be optimized. Let us consider the two-dimensional design space,
discretized by one of two different methods. The first method to employ is the grounded
structure approach (GSA). GSA divides a workspace up into nodes and all of the elements that
connect those nodes, as in Figure 18. The design space is the interconnected network of beams,
from which the structure and actuators of the final system will be selected. GSA has been widely
implemented in compliant mechanism and structural optimization and is chosen for this research
because the structural form of actuators tends to be beam-like in shape and connectivity. One of
the drawbacks of using GSA is that structural connectivity of the inputs, outputs, and ground is
not guaranteed. Solutions to this issue include graph-based constraints that check for valid
structures.
As mentioned in the Background section, another method of discretization has been
developed: the Load Path Method (Lu, 2004). The load path method begins with connected
graph components, ensuring connectivity of all design configurations. In the load path method,
key parts of the design domain are identified, such as the input port, the output port, and the
fixed ports. Variable “load paths” are established between these ports, which will eventually
form the final structure, guaranteed to be connected.
Originally developed to synthesize shape-morphing compliant mechanisms, the load-path
representation (Lu, 2004) is a design domain parameterization that represents the structural
topology in terms of the physical connections between the input and the output points. These
physical connections are the load-paths through which the energy from the input actuator can be
delivered to output points. The structural topology is determined by the presence/absence of
each path as well as the connectivity between different paths. A binary topology variable is
assigned to each load-path to represent the presence of the path, while a set of intermediate
connection ports controls the connectivity between different paths. This representation is
incorporated into a genetic algorithm to synthesize shape-morphing compliant mechanisms.
Figure 32 shows an example topology using the load-path representation. The design
domain is reduced to only the upper half of the triangular area to achieve symmetry. Five
connection ports are used in this design to control the intermediate connections between different
paths. The locations of the connection ports may also wander within the design domain to create
Brian Trease, University of Michigan Page 40 of 63 6/9/2005
various topologies. Figure 33 shows how the topology changes according to the binary topology
variables. The connection port locations are the same in Figure 32 and Figure 33, but the binary
topology variables listed in Table 4 have different values (e.g. paths 1, 4, and 13). Furthermore,
the locations of the intermediate connection ports (e.g. node 11) can also wander within the
design domain, leading to yet another different topology, shown in Figure 34.
Figure 32: Example topology for Load-Path method
Figure 33: Effect of changing binary topology variables
Brian Trease, University of Michigan Page 41 of 63 6/9/2005
Table 4: Topology Variables used in Figure 32 and Figure 33 Path type
Path No.
Path sequence Figure 18 Figure 19
1 {1,9,4} 1 0 2 {1,5} 1 1 3 {1,10,6} 0 0
in ∞
out 4 {1,11,9,7} 1 0 5 {1,11,2} 1 1 in
∞ sym 6 {1,8,3} 0 0
7 {2,9,4} 0 0 8 {2,5} 0 0 9 {2,8,6} 0 0 10 {2,7} 0 0 11 {3,4} 0 0 12 {3,9,5} 0 0 13 {3,11,6} 0 1
sym ∞
out
14 {3,8,7} 0 0
Figure 34: Effect of wandering connection port
It may also be of interest to investigate other graph-based methods for mapping the
design space. These include permutating a basic graph that already has the minimal connectivity
requirements. Another option is to conduct graph searches over an entire Grounded Structure
network. In the latter, individual elements would no longer be the variables, but rather their
connected graphs.
Actuator models serve as both force generators and structural elements. Actuators must
be properly parameterized to include their various force-deflection responses in both the
activated and inactivated states. These constraints can all be implemented numerically within the
optimization software. In both single and multiple output scenarios, every element has the
Brian Trease, University of Michigan Page 42 of 63 6/9/2005
discrete potential to be one of the following: an actuator, a structural elastic member to support
loads and undergo controlled deformation, or absent from the design.
8.1.3. Nonlinear Topology Optimization and Synthesis Given a model of the design space, all the variables can be optimized via a genetic
algorithm (GA) to yield a mechanism topology. GA’s are a form of nonlinear optimization that
find global optima as opposed to local optima. Genetic algorithms convert all the design
variables into a long sequence of genes that can be propagated from generation to generation
with the effects of cross-over and mutation. The best-fit individuals survive by scoring high
against an objective (fitness) function. The GA starts with a population of randomly generated
designs. The selection scheme in GA is based on the ‘survival of the fittest.’ The performance
of each design is evaluated using the objective function. Once the global part of the algorithm
finds the basin of convergence of the optimum, the local part of the algorithm (size/shape
optimization) quickly and automatically exploits it.
While the optimizer is nonlinear, the structural analysis within the optimizer may be
linear or nonlinear. For more information refer to the Objective Functions section (8.1.4).
Genetic algorithms are chosen for several reasons. GA’s are well suited to handle the
discrete nature of the problem of actuator and element distribution. Elements either must exist or
be removed; there is no choice in between. Further, previous genetic based algorithms for
compliant mechanism synthesis have outperformed the continuous optimization methods (Lu,
2004).
To aide in convergence of the algorithm, both Darwinian and Lamarckian evolution are
employed. Lamarckian evolution refers to the actual modification of the genome (genetic
engineering) before it is passed on to offspring. The motivation and details of this are described
in the Preliminary Work and Results section (7.5). When modifications occur before fitness
function assessment, the term genetic engineering applies. Modifications after assessment are
based on performance and referred to as Lamarckian engineering.
While GA’s are very powerful for finding global optimums in difficult, nonlinear
problems, they suffer from high computational cost. I will also examine other nonlinear
optimization methods, such as stochastic local search (Hoos, 2004).
Brian Trease, University of Michigan Page 43 of 63 6/9/2005
8.1.4. Objective Functions A key aspect of the research is implementing the proper objective function to achieve
meaningful results. The choice of objective function depends on the type of problem being
solved, which has already been described (SISO, MISO, MIMO, shape-change, etc.) The
objectives include energy efficiency with regards to input work and output work. The
maximization of energy efficiency may lead to distributed compliance in our structures. Other
objective functions may be applied to minimize actuator power consumption, minimize the total
number of actuators used, or to maximize the combined effect of multiple actuators. Shape
change objectives have been implemented via least squares formulations, by calculating the
deviations of points on a deflected structure from corresponding points on a target curve.
The quantities used in the various objective functions will be calculated using beam
elements within a linear finite element analysis. The analysis will be written in Matlab. Should
synthesis results indicate a need for nonlinear analysis within the optimizer, FEMlab, a
commercial Matlab toolbox, will be used to reduce run-times.
The design objective functions will be developed with increasing levels of complexity in
design goals, working our way to the fully embodied problem, as follows:
(i) maximize energy efficiency of the embedded actuator compliant mechanisms, plus
(ii) ⊕ optimal layout of sensors, plus
(iii) ⊕ controllability including ‘orthogonality’ of ‘deformation modes’, plus
(iv) ⊕ optimal layout for distributed sensing ‘internal stress state’, plus
(v) ⊕ multiple load-case objectives for dynamic/adaptive performance
8.1.5. Constraints In seeking various objectives that will bring us toward an optimal system, several
practical constraints will also be placed on the models such as weight, energy efficiency, peak-
power consumption (especially important for autonomous robots), materials (permissible stresses
and strains), and manufacturing (minimum width and thickness), dynamic performance
(responsiveness, adaptability, variable stiffness, natural frequency), and the type of actuators
(force-displacement curves). Of particular importance are constraints that guarantee the ground,
input, and output are all connected to the same structure, and that there are no other floating
structures. Additional research is required to identify the best means of ensuring structure
connectivity. Constraints are needed to guarantee specified output displacement, while keeping
Brian Trease, University of Michigan Page 44 of 63 6/9/2005
deflection in undesired directions below a specified limit. Stress constraints are also important,
and will be applied later in Size/Shape Optimization.
8.1.6. Cross-over strategies It is desirable to have a meaningful cross-over strategy that corresponds to the actual
problem when devising genetic algorithms. That is, the exchange of genes should correspond to
the exchange of subsystems in the phenotype. (i.e. genotype sequencing does not map to
phenotype expression.) While the Grounded Structure Approach makes this difficult, either an
appropriate strategy will be sought, or another discretization will be employed (i.e. Load Path
Method.)
8.1.7. Genetic Engineering I propose to apply “genetic engineering” to our design population members during the
GA. This means trimming away substructures that are not affecting the operation of the
mechanism. Graph search is used to find elements that branch off the main structure and do not
connect to anything else. Elements that simply undergo rigid body motion do not contribute to
the structural function and are also eliminated. The designs are effectively “cleaned up” and
only the core elements are passed on to future generations. For further details please refer back
to the Preliminary Work and Results: Modified Genetic Reproduction section (7.5).
8.1.8. Size / Shape Optimization It is typical to separate the topology optimization from the size and shape optimization in
the synthesis of compliant mechanisms. Topology optimization for the current problem has
already been described; it is the discrete choice of the elements that create the unique physical
layout of the structure to satisfy kinematic motion or shape-morphing function. This establishes
a functional design. However, many other practical constraints and performance requirements
must be met before the design is deemed useful. These include materials (permissible stresses
and strains), desired fatigue life, prevention of localized buckling, natural frequency, making
sure that the dimensions and shapes can be physically realized (manufacturing constraints), etc.
All these constraints are accounted for during size/shape and geometry optimization. Fine tuning
of the mechanical function can be achieved (such as mechanical or geometric advantage) along
with the monitoring and control of stress constraints.
Brian Trease, University of Michigan Page 45 of 63 6/9/2005
During this size optimization, the nodes or interconnection points are allowed to wander
within a certain window thereby modifying the geometry without altering the topology. Size and
shape optimization is performed on an already fixed topology, only changing the node locations
and beam width and thickness. The objective function might be stated as maximizing energy
efficiency of the system, maximizing the energy storage, or reduction of the peak force or peak
power consumption, depending on the nature of the task. Figure 7 shows a schematic of how
various dimensions, geometry, and shapes are optimized during size optimization. The results of
Stage I and Stage II have the same topology but different size and shape.
8.1.9. Finite Element Confirmation To confirm that the genetic algorithms generate feasible structures that actually display
the desired attributes, the resulting structures will be modeled and analyzed with a commercial
finite element code such as ANSYS. Rather than beam elements, solid elements will be used,
allowing for the inclusion of stress concentration effects.
8.1.10. Physical Prototype In addition to the finite element analysis, a physical prototype will be constructed to
demonstrate proof of concept.
8.1.11. Integration of Supplementary Concepts While the synthesis of multiple actuator compliant systems is the main focus, there are
many other ideas I am interested in exploring and integrating within the design optimization as
extensions to the basic synthesis framework.
8.1.11.1. Optimal sensor layout Traditional mechanisms allow motion at a discrete set of sliding or rolling interfaces, e.g.
bearings. Appropriate motion sensors, such as potentiometers, LVDTs, and optical encoders, are
situated on the translational or rotational axes. They are discrete sensors, matched to the discrete
motions of traditional mechanisms.
Compliant mechanisms are dramatically different, allowing motion by distributed flexing
of continuous media. As an analogue to optimal actuator layout, I am also interested in optimal
sensor layout. As with actuators, we can also speak of embedded and distributed sensing. Work
from our peers and collaborators at NWU has resulted in the development of Mutual Inductance
Brian Trease, University of Michigan Page 46 of 63 6/9/2005
Sensing. They developed a distributed sensing technology that is well matched to the distributed
motion of compliant mechanisms. The technology consists of embedded wires (or other
conductive traces) in the compliant members. A pair of such traces – one may be thought of as a
transmitter and one as a receiver – communicate electromagnetically, coupled by their mutual
inductance, which varies in response to deflection of the compliant member. Importantly, the
signal from a distributed sensor is an integrated signal, which accumulates the deflection of the
compliant member along its whole extent. Such sensors readily lend themselves to the compliant
mechanism framework, as has already been physically demonstrated by collaboration with UM
and NWU.
A key feature of the technology is that the transmitters and receivers may consist of
single traces (multi-turn coils are not necessary.) The resulting signals are small but easily
detected with modern electronics. Single traces can be fabricated inside or applied to the surface
of compliant members. On traditional mechanisms, signals from distributed sensors would need
wires to cross the gaps at other axes of motion, but on compliant mechanisms there need not be
such gaps. In some uses, there may be many receivers for a single transmitter. An example of
the use of distributed sensors on a compliant mechanism is measuring the deflection and shape of
a compliant wing or fin, in an autonomous aerial or underwater vehicle.
Further integration of these distributed sensors within a compliant systems environment
will require continued research into suitable fabrication methods for embedding them in
compliant mechanisms. Our collaborators continue efforts to miniaturize the sensor electronics
so that a significant density of sensors can be incorporated into a compliant mechanism
efficiently. In the meantime, I focus on a general theory and method for inclusion of sensors in
compliant systems.
8.1.11.2. Design for Control In what the literature indicates to be a first, I also seek to implement “design for control”
within our optimization strategies. The step from single-actuator to multiple-actuators in CM
design begets the question of how to properly control these multiple actuators. With the sensor
networks described above, I will explore the design of an embedded actuation system capable of
responding to varying and unpredictable environments. While it is possible to add sensors and a
controller to an already finished multi-actuator compliant device, I believe these designs can be
vastly improved by considering sensors and controls during the layout of the compliant structure.
Brian Trease, University of Michigan Page 47 of 63 6/9/2005
8.1.11.2.1. Internal stress state sensing / correspondence to external loadings More specifically, I seek to implement effective control that uses internal stress state
sensing to correspond to external loads. To capture this intent within the optimizer, we borrow
the terms "controllability" and "observability" from the field of control theory. A task of this
research is to create specific, mathematical definitions of these terms as they apply to spatial
inputs and outputs of compliant structures. While it may result that our definitions are only
analogous to those from controls, it may turn out the specific parts of the controls definition
apply directly to our problem. From Ogata’s Modern Control Engineering textbook, these terms
are defined:
Controllability: A system is said to be controllable if it is possible by means of an unconstrained control vector to transfer the system from any initial state to any other state in a finite interval of time.
Observability: A system is said to be completely observable if every state can be determined from the observation of y(t) over a finite time interval. The system is, therefore, completely observable if every transition of the state eventually affects every element of the output vector. This is useful in reconstructing unmeasurable states from measurable ones.
I propose to explore the internal stress states as a means to map the external conditions.
This might be particularly useful, for instance, when the on-set of stagnation pressure on an
airfoil must be sensed slightly in advance to signal an appropriate actuator to deform the trailing
edge by a certain amount, thereby changing the angle of attack. Certain internal elements of the
compliant system will be designed to deform predictably to maximize the sensitivity of a sensor
embedded within these elements. Although force can be sensed directly at the point of interest
with pressure sensors, the external pressure forces could also be interpolated from internal stress
state of structure. Utilization of such internal sensors has the potential to measure “modes” of
pressure deflection. Summing various modes of deflection may be better than directly measuring
along the surface of interest. This will further protect the sensors from the environment and
fewer internal sensors may be needed. More importantly, the placement of internal sensors
avoids the tradeoff of needing the output to be stiff against external loads yet compliant enough
to create measurements.
The topology synthesis problem will have to be sophisticated enough to capture the
knowledge of the controller into the topology synthesis problem formulation. For instance,
Brian Trease, University of Michigan Page 48 of 63 6/9/2005
additional constraints will be needed to ensure ‘orthogonal deformation modes’ for the actuators
and ‘orthogonal sensing modes’ for the sensors.
8.1.11.2.2. Implement knowledge of the controller in the initial actuator/structure/topology synthesis
With a method for calculating these terms for a given compliant device, one can include
these terms within the objective function or constraints of the optimization. I hypothesize that
linear independence (orthogonality) of the deformation shapes (modes) created by each actuator
will lead to a greater degree of freedom for controlling external aspects of the structure.
Likewise for sensors, they can be arranged for maximal sensitivity to a variety of external
loadings of the structure. The question remains as to whether these are better posed as objectives
or constraints.
8.1.11.2.3. Effective control methods As well as component layout for optimal control, one should also consider the most
effective control methods for actuators and their compliant transmissions, taking advantage of
feedback control. This is complicated by the fact that in compliant mechanisms essentially every
closed-loop control problem is distributed control; we do not have lumped-parameter systems.
Yet, it ought to be possible to design the embedded sensor such that its distributed sensitivity can
simplify the distributed actuation control problem.
8.1.11.3. Variable stiffness concept Variable stiffness of compliant mechanisms is an interesting open research question.
With regards to this proposal, I see variable stiffness as one of the biological imperatives to
explore, which is further motivated in the Biomimetics section. However, as my own and
outsider interest continues to grow, the appeal of mathematically studying this problem and
implementing it in our optimization becomes stronger.
In one sense, variable stiffness is another way to approach some of the problems that are
directly handled by the current objective formulations. Rather than desiring to resist an external
load directly, via an actuator and a leverage scheme, we can instead resist the external load by
changing the stiffness of the structure to meet it. This change is caused by actuators, which alter
the geometry to make the structure stiffer. A simple demonstration of this is changing the
Brian Trease, University of Michigan Page 49 of 63 6/9/2005
orientation of beams to be in line with external forces, exploiting the axial stiffness of the beams,
which is much greater than their bending stiffness.
8.2. Actuation Concurrent with the development of the above design methodology, I will continue to
study the best actuator technology to integrate with our systems. Thus, one aspect of our design
methodology will be actuator parameterization, so that I can create a general embedded actuation
system methodology for which any particular actuator's characteristics only need be specified. I
have experience in taxonomy of actuator technologies; Figure 35 compares a wide variety of
actuator types (Trease, 2001).
Most actuators can be characterized by four key design specifications: force, size,
displacement, and frequency. Power consumption can be treated as a cost factor for further
comparison. Typically, actuators with high displacement provide low force and vice versa.
Even for a given actuator, the maximum force and maximum displacement are not achieved at
the same time. The maximum force is the “blocking force” measured when there is zero
displacement. The maximum displacement is the “free displacement” measured when there is no
load in resistance. Therefore, each actuator is not represented by a point on this graph, but by a
curve. It is therefore critical that actuator selection is based on the entire motion characteristics
(force displacement curves) rather than single-point designs. Ashby (1997) provides a
systematic method of selecting actuators.
Brian Trease, University of Michigan Page 50 of 63 6/9/2005
Figure 35: A map of the actuator technology workspace, with respect to Force, Displacement, and Work.
B. Trease, 2001, A Survey and Comparison of Smart Material Linear Actuators
8.2.1. Actuation Selection and Evaluation Actuators best matched to our needs will be those that exhibit capability as structural
elements. Actuators of interest include Electrostrictive Polymer Artificial Muscles, Electrically-
activated PAN muscles, Ionic Metal Composite muscles, and Shape-Memory-Alloy (SMA)
actuators. Studies can evaluate which actuators are the most practical, and how their limitations
affect the research objective. For example, shape memory alloy wires on their own provide little
structural stiffness except in tension. Piezoelectric benders may serve well as long as not loaded
in tension. A possible solution is smart actuator composites: a composite embedded with shape
memory alloy wires would be able to support bending and compressive loads.
8.2.2. Characterization As stated before, I desire to implement any actuators in the code as more than mere point
forces or displacements. Our approach will focus on several increases in actuator model fidelity.
Brian Trease, University of Michigan Page 51 of 63 6/9/2005
Since the simplest model is a mere point force, the next step is to add a structural element with
the stiffness of the actuator. However, this case still implies that an actuator can provide a
specified force or displacement, which in reality requires the use of closed-loop feedback
control. A powered actuator generates a force that corresponds to its unique load-displacement
actuation curve and the stiffness of the system on which upon it acts. Thus, we need to capture
this load-displacement curve within our finite element calculations. The FEA solution will
determine the actual actuator force and displacement.
In addition, I may take a solid continua approach to the mathematical modeling of these
actuators, developing 3-D stiffness matrices and coding the nonlinear force-deflection
relationships.
8.2.3. Commercial Artificial Muscle Actuators While I will consider all types of actuator technology, of particular interest are
Commercial Artificial Muscle Actuators. These are among the cutting edge of actuator
technology, often supplying the greatest displacements and strains compared to others. Their
biological-inspiration may indicate suitability for use in the proposed research. However,
research must be done to ensure that the match is genuine and not merely based on the word
“biomimetic.”
Figure 36: Commercial Artificial Muscle Actuators. Bending and linear types are shown.
http://www.artificialmuscle.com/
8.3. Bio-Discussion – “Biomimetics”
8.3.1. Observations & Comparisons “Biomimetic” can be an ambiguous term and this research considers its useful meaning it
at least three different ways. First, and most simple, is the observation that our synthesis results
from and compares to structures from biological systems. For example, proper implementation
of the methodology will allow the following question to be answered: are optimal
Brian Trease, University of Michigan Page 52 of 63 6/9/2005
structure/actuator distributions similar to the musculo-skeletal systems found in nature? This is
not a guaranteed form of scientific inquiry, yet nature may provide some inspiration for one to
interpret what we have created. Existence of such evidence, or lack thereof, may support our
claims as to the strength of our methodology.
8.3.2. Inspiration via Direct Borrowing Next, as perhaps as most commonly done, we can think of biomimetics in terms of direct
borrowing from nature. This implies observation and study of specific structures in nature and
directly implementing them in engineered designs. While this is an exciting area of research,
nature is neither guaranteed to provide optimal designs nor designs appropriate for emulation
with technology. One can think of many man-made designs that surpass nature, from the
combustion engine to fixed-wing aircraft.
Figure 37: Spawlita, from the Center for Design Research at Stanford
Some examples of direct borrowing include various bio-robot and insectoid-robot
projects, such a Spawlita, the hexapedel biorobot shown in Figure 37. Compliant mechanisms
are of immediate appeal in addressing the problem of complex, biomimetic deformation. A
fully-compliant system also has great stealth potential for both aquatic and aeronautical craft.
Smooth, continuous deformations result in a much less distinguished sonar “signature”, but even
if spotted, biomimetic robots/vessels may be mistaken for their biological precedents. I have
Brian Trease, University of Michigan Page 53 of 63 6/9/2005
worked on such projects, including a swimming fish project. This project sought to mimic the
kinematics of biological fish stroke data, while also using a compliant skeleton inspired by that
of a fish fin (Trease and Lu, 2003).
Yet in the end, the question remains: why force the issue? If specific structures are
indeed best, then they will also result from a properly design optimization, and we need not have
borrowed them directly.
8.3.3. Inspiration via Objective / Task Another form of biomimicry is inspiration via objective or task, which is an indirect
means of emulating nature. In other words, we should try to mimic the apparent goals and
purposes of nature, rather than nature directly. The question of what exactly from nature is
worthy of using as an objective function is interesting in itself. Some possibilities have already
been mentioned, such as variable stiffness and internalized sensing and actuation. Variable
stiffness is often employed in nature to alter passive dynamics, and is further discussed in the
Case Studies section. The concept of internal structures interacting with the external
environment was described in terms of distal attribution in the Vision section of this proposal.
Biology offers the additional case of embedded sensing for closed-loop muscle control. Pacinian
Corpuscles are elements of the haptic feedback system of human touch. They are structures
found within the skin that measure pressure and vibration so that forces may be estimated for
feedback in control (Thompson, 2001). Indeed, the idea of closed-loop force control is an area
from which to further seek biological objectives for implementation in our methods.
8.3.3.1. Homeostasis / Homeokinesis Along the lines of the internal actuation concept, the objective of embedded compliant
systems may be described as ‘homeostasis” or “homeokinesis”. These terms refer to internal
processes in biology that attempt to maintain internal balance. As the proposed structures can be
viewed as internally driven black boxes converting inputs into outputs, homeostasis refers to the
maintenance of a stable state and homeokinesis refers to such maintenance in spite of changing
environmental conditions.
8.3.3.2. Energetics Finally, the energetics of biological functions may also provide additional objectives. We
currently seek to define energy efficiency in a number of ways, such as work output divided by
Brian Trease, University of Michigan Page 54 of 63 6/9/2005
work input, work throughput, and mutual strain energy. Nature may provide other formulations
of energy efficiency. For example, energy efficiency may be viewed in terms of the energy
potentially saved due to passive dynamics and the use of variable stiffness.
8.3.4. Genetic Algorithms Our choice of genetic algorithms to solve biologically-inspired problems may appear to
enable synergetic benefits, but this must studied and confirmed with stronger argument. As with
the use of artificial muscle actuators, for now it is only coincidence I am using biomimetic
algorithms to solve a biomimetic problem.
8.3.5. Idea of an integrated Musculo-Skeletal-Ligament-(Nervous) System Some background information about the idea of an integrated musculo-skeletal-ligament-
(nervous) system reveals the following views. Though it is distinctly a view with a lot of
potential, the idea of such integration at the system level is not a clearly articulated common
view of many in biology. Yet, it is a logical viewpoint from a design point-of-view. In fact, its
analogy in Mechanical Design is the defined goal of my proposed research. Not the mimicry of
specific structure, but the analogy of a combined structure-actuator-sensor system is the general
concept.
8.3.6. Biological Evidence of Distributed Actuation As an important note of the appropriateness of biomimicry, discussed here is biological
evidence of distributed actuation. This, again, refers to systems where the actuators/muscles are
embedded within the moving structures. There are questions among biologists whether such
actuators are used mainly in attenuation and shape-control situations. (e.g. camber control of
fish fins.) It is not as evident that embedded muscles are as useful for power generation or thrust.
Larger muscles are typically outside of the system due to their bulkiness, and attached to part of
the external skeleton. This debate, however, depends or how the system is defined; for some
muscles are clearly part of the system, such as those in the lower leg.
On the other hand, this may be the case where we look at nature and surpass it. Our
engineered designs have advantages that nature does not, such as high power density actuators,
which may enable force generation via embedded actuators.
Brian Trease, University of Michigan Page 55 of 63 6/9/2005
9. Confirmation of Results The initial results of the genetic algorithms should be verified with commercial finite
element in ANSYS. Rapid-prototyping technology will be used to create initial physical models.
These models will serve not only as geometry checks, but can also be used for performance
testing. The next step in fabrication is investigation and sourcing of multi-material fabrication
for monolithic structures.
10. Case Studies Several case studies will be performed to evaluate the quality of the topology synthesis.
These include a variable stiffness problem, an adaptive orthosis/prosthesis problem, and an
adaptive shape-changing wing problem.
10.1. Variable Stiffness As already mentioned, there are many examples where nature uses variable stiffness in
structures. It is often used for adjusting the passive dynamics of a system. For example, many
animals will modify their stiffness so that their natural frequency matches their walking or
running speed. Variable stiffness is also used to prepare for impact absorption or sudden
movements.
Brian Trease, University of Michigan Page 56 of 63 6/9/2005
equi
vale
nt s
tiffn
ess
com
plia
nt m
echa
nism
unactuated and compliant actuated and stiff
ground
output
Figure 38: Demonstration of Variable Stiffness achieved by Manipulation of Geometry
At least three means of achieving variable stiffness can be identified from nature and
possibly implemented as optimization fitness functions. First is the case of direct and active
force application to balance the external forces and result in an apparent change in stiffness. In
animals this is often achieved by co-contraction of opposing muscle groups. For example,
visualize a chef slicing through a stick of butter with an extended arm and then stiffening up the
arm to cut through an adjacent piece of thick steak.
Another way to change stiffness is to alter the structural geometry of the system (Figure
38). Using actuators or muscles to rearrange elements can change the effective load paths or
cause elements to use axial modes instead of bending modes and vice versa.
Finally, some muscles can change their very material stiffness to directly achieve this
effect. Similarly, there are now new smart materials that can change their stiffness under applied
electric or magnetic loads. While this is beyond the scope of the proposed work, the mechanism
behind changing muscle stiffness may be studied for further insight.
Brian Trease, University of Michigan Page 57 of 63 6/9/2005
10.2. Adaptive Orthotic and Prosthetic Devices Another case study will be the design of an adaptive orthosis. Many current devices must
be custom fit to every individual and then cannot be changed during use. A self-adjusting device
would resolve both of these issues. A person’s limb will swell and change shape during use,
creating discomfort for the user. With a system capable of sensing the external pressure load and
changing the shape to minimize any stress concentrations, discomfort can be alleviated. Such
devices could also be used for functional changes in different scenarios. For example, an
orthosis could adjust to increase stiffness during running and before jumping.
10.3. Shape-changing wings and fins A third case study is a continuation of a standard compliant mechanism problem, the
shape-changing aircraft wing or aqua-craft fin. In this study, however, the goal would be to
create as system that could sense a variety of external pressure load profiles and respond with a
variety of shape changes. Such sense-and-control variable geometry wings could adjust to the
most appropriate and efficient configuration for both predicted and unpredicted flight conditions.
These benefits would also apply to loading-dependent shape-changing aqua-craft fins.
11. Research Timeline ● = deadline
X = conference Jun-Aug Sep-Nov Dec-Feb Mar-May Jun-Aug 2006
Embedded Actuator Synthesis
Load Path Approach Actuator Stiffness
Properties Controls Groundwork Biomimetic Studies
Biomimetic Implementations
Prototyping Diss. Writing
Res
earc
h P
rogr
ess
Journal of Mech. Design ●
SPIE Conf. ● X ASME Design Conf. ● >X
ASME IMECE Congress ● >XDea
dlin
es
ASME/IEEE Mechatronics ●
Brian Trease, University of Michigan Page 58 of 63 6/9/2005
12. Potential Contributions The main potential contributions of the proposed research are as follows:
• A method for including actuators as variables in the optimization will be developed. The number, location, and orientation of the actuators will all be variables subject to change.
• This research will establish the benefit of including actuator stiffness and performance characteristics within the structural synthesis.
• Another contribution will be the consideration of “design for control” within compliant mechanisms.
• Finally, I aim to discover new aspects of the relationship between nature and design, addressing the appropriate use of biomimicry.
The key project deliverables will be mathematical characterizations, automatic design
synthesis software, and two journal papers tentatively on the following topics:
• Theory and Design Methodology for Synthesis of Compliant Mechanisms with Embedded Actuation
• Design for Embedded Actuation, Sensing, and Control in Distributed Compliant Systems
In addition, I propose a third journal article to write after September 2006 and completion
of the doctoral degree:
• Biologically-inspired Design and Interpretation of Adaptive, Distributed, and Embedded Compliant Systems
The proposed research advances the current state of the electromechanical systems design
in many fundamental ways. Traditionally, structures, mechanisms, actuators, and sensor design
is the focus of different people paying marginal attention to one another, resulting in suboptimal
systems. The proposed research paves way to a scientific approach to true mechatronic system
design rather than focus primarily on structures, sensors, or actuators.
The result of the proposed biologically-inspired research is a fundamental understanding
of what it takes to generate a globally-optimal solution for compliant systems with embedded
actuation and sensing. In sharp contrast to the current practice of “slapping on” actuators and
sensors to a finished design, the creative stage of design will simultaneously consider the
structural stiffness, power consumption, and actuator and sensor characteristics.
Brian Trease, University of Michigan Page 59 of 63 6/9/2005
13. Summary In conclusion, I am leveraging my previous research experiences to develop a new design
paradigm for creating biologically-inspired devices for a variety of applications. I will introduce
many new aspects to the field of compliant mechanism design, including embedded actuators,
design for control, and genetic engineering within the synthesis algorithms. While pursuing this
goal, I while pay special attention to the methods and benefits of applying biomimicry to
mechanical system design. The culmination of my Ph.D. work will be transforming computer-
generated designs with optimized geometry, materials, and properties directly into the physical
realization of functional, embedded, and distributed compliant systems.
Brian Trease, University of Michigan Page 60 of 63 6/9/2005
14. References
14.1. General Huber, JE, Fleck, NA and Ashby, MF. (1997). “The selection of mechanical actuators based on performance indices.” Proc R. Soc. 453: 2185-2205. Holger H. Hoos and Thomas Stützle: Stochastic Local Search: Foundations and Applications. Morgan Kaufmann, San Francisco (CA), USA, 2004 Ogata, Katsuhiko. Modern Control Engineering, Englewood Cliffs, NJ: Prentice Hall, 1990. Trease, B. P., 2001, “A Survey and Comparison of Smart Material Linear Actuators,” University of Michigan / WPAFB Internal Report, www-personal.engin.umich.edu/~btrease/share/Trease_Actuator_Report.doc
14.2. Compliant Mechanisms Anathasuresh, G.K., “A New Design Paradigm for Micro-Electro-Mechanical Systems & Investigations on the Compliant Mechanism Synthesis”, PhD Dissertation, Department of Mechanical Engineering, University of Michigan, Ann Arbor. Frecker, M.I., Ananthasuresh, G.K., Nishiwaki, S., Kikuchi, N., and Kota, S., 1997, "Topological Synthesis of Compliant Mechanisms Using Multi-Criteria Optimization," ASME Journal of Mechanical Design, 119(2):238-245. Hetrick, J.A., 1999, "An Energy Efficiency Approach for Unified Topological and Dimensional Synthesis of Compliant Mechanisms," Ph.D. Dissertation, Department of Mechanical Engineering, University of Michigan, Ann Arbor. Hetrick, J. and Kota, S., 1999, "An Energy Formulation for Parametric Size and Shape Optimization of Compliant Mechanisms," ASME Journal of Mechanical Design, 121:229-234. Paros, J.M. and Weisbord, L., 1965, “How to Design Flexure Hinges”, Machine Design, pp. 151-156 Joo, J., 2001, "Nonlinear Synthesis of Compliant Mechanisms: Topology and Size and Shape Design," PhD Dissertation, Mechanical Engineering, University of Michigan, Ann Arbor. Joo, J., Kota, S., and Kikuchi, N., 2000, "Topological Synthesis of Compliant Mechanisms Using Linear Beam Elements," Mechanics Based Design of Structures and Machines, 28(4):245-280. Joo, J., Kota, S., and Kikuchi, N., 2001, "Large Deformation Behavior of Compliant Mechanisms," ASME 2001 Design Engineering Technical Conference, Pittsburg, PA, DETC2001:DAC-21084.
Brian Trease, University of Michigan Page 61 of 63 6/9/2005
Kota, S., Rodgers, M.S., and Hetrick, J.A., 2001, "Compliant Displacement-Multiplying Apparatus for Microelectromechanical Systems," United States Patent No. 6,175,170. Hetrick, J. and Kota, S., 2003, "Displacement Amplification Structure and Device," United States Patent No. 6,557,436. Howell, L.L., 2001, "Compliant Mechanisms," John Wiley and Sons, Inc. Saxena, A. and Ananthasuresh, G.K., 2001, "Topology Optimization of Compliant Mechanisms with Strength Considertations," Mechanics of Structures and Machines, 29(2):199-221. Saggere, L., 1997, "Static Shape Control of Smart Structures: A New Approach Utilizing Compliant Mechanisms," Ph.D. Dissertation, Department of Mechanical Engineering, University of Michigan, Ann Arbor.
14.3. Related Work Anusonti-Inthra, P., Gandhi, F. , and M. Frecker, 2003, Design of a Conformable Rotor Airfoil Using Distributed Piezoelectric Actuation, Proceedings ASME International Mechanical Engineering Congress and Exposition, Adaptive Structures Symposium, Washington, DC, November 16-21, 2003. Paper IMECE2003-42659. Bharti, S., and M. Frecker, 2003, Compliant Mechanical Amplifier Design Using Multiple Optimally Placed Actuators, Proceedings ASME International Mechanical Engineering Congress and Exposition, Adaptive Structures Symposium, Washington, DC, November 16-21, 2003. Paper IMECE2003-42658. Silva, E.C.N., 1998, “Design of Piezocomposite Materials and Piezoelectric Transducers using Topology Optimization”, PhD Dissertation, Department of Mechanical Engineering, University of Michigan, Ann Arbor.
14.4. Load Path Method Akhtar, S., Tai, K., and Prasad, J., 2002, "Topology Optimization of Compliant Mechanisms Using Evolutionary Algorithm with Design Geometry Encoded as a Graph," ASME 2002 Design Engineering Technical Conferences, Montreal, Canada, DETC2002:DAC-34147. Lu, K., 2004, “Synthesis of Shape Morphing Compliant Mechanisms”, PhD Dissertation, Department of Mechanical Engineering, University of Michigan, Ann Arbor. Tai, K. and Chee, T.H., 2000, "Design of Structures and Compliant Mechanisms by Evolutionary Optimization of Morphological Representations of Topology," ASME Journal of Mechanical Design, 122:560-566. Tai, K., Cui, G.Y., and Ray, T., 2002, "Design Synthesis of Path Generating Compliant Mechanisms by Evolutionary Optimization of Topology and Shape," Journal of Mechanical Design, Transactions of the ASME, 124(3):492-500.
Brian Trease, University of Michigan Page 62 of 63 6/9/2005
14.5. Biology Loomis, J. M. (1992) Distal attribution and presence. Presence: Teleoperators and Virtual Environments, 1, 113-119.
Thompson Richard Lee. "Integration of Visual and Haptic Feedback for Teleoperation". www.robots.ox.ac.uk/ActiveVision/ Papers/thompson_dphil2001/thompson_dphil2001.pdf Trease, B.P., Lu, K.J., and Kota, S., Biomimetic Compliant System for Smart Actuator-Driven Aquatic Propulsion: Preliminary Results, ASME International Mechanical Engineering Congress & Exposition, Washington, D.C., IMECE2003-41446, November 16-21, 2003
Brian Trease, University of Michigan Page 63 of 63 6/9/2005