Statement of Research and Teaching Interests Research Statement

Jyh-Ming Lien/∼neilien/Statement of Research and Teaching InterestsResearch StatementMy research interests are in the areas of computational geometry and computer graphics.A theme of my research is to establish practical algorithms for representing,computing,and approximating shapes and motion in geometric computations.My research has contributed to the development of algorithms for real-world problems in a wide range of application domains including robotic motion planning [18,20,23],agent-based simulation [1,14,16,17],artificial life [2,5],computer animation [3,4,6],geometric processing [7–13],human-computer interface [21,22],computational neuroscience [15]and computational geoscience.The best way to illustrate of my work is through the graphics and animations that can be found on my web page.Shape and motion are the essence of many geometric problems.We have developed techniques for ap-proximating shape and motion in a range of applications that provide computational efficiency and also simultaneously allow multi-resolution or hierarchical representations.Additionally,in many cases the ap-proximate representations facilitate the identification of “significant”structural features and enable us to ignore less significant features that are not important to the problem at hand.In the following,we briefly describe some of our work in approximating and representing shape,generating and controlling motion,and generating and simulating group behaviors.Techniques for Approximating and Representing ShapeApproximate Convex Decomposition (acd ).Convex decomposition can be useful because many problems can be solved more efficiently for convex objects.However,generating convex decompositions can be time consuming (sometimes intractable)and can result in unmanageably largedecompositions.(a)(b)Figure 1:(a)An exact convex decomposition and an acd (with convexity less than 0.04)of the David model have 85,132and 66components,resp.(b)The convex hulls of the acd components represent David’s shape.To address these issues,we propose a partitioningstrategy that decomposes a given 2D or 3D modelinto approximately convex components,resulting inan approximate convex decomposition (acd )[7–11].We compute an acd of a model recursively until allcomponents in the decomposition have concavity lessthan some specified (tunable)parameter.Examplesof acd are shown in Figure 1.For many applications,the approximately convexcomponents of our acd provide similar benefits asconvex components,while the resulting decomposi-tion is both significantly smaller and can be computedmore efficiently.We have shown both theoreticallyand experimentally that the acd of polygons withzero or more holes and polyhedra with arbitrary genus can efficiently produce high quality decompositions.Applications that can benefit from this approach in-clude collision detection [11],penetration depth esti-mation,mesh generation [19],and motion planning.Another important aspect of an approximate con-vex decomposition is that it can more accurately rep-resent important structural features of the model byproviding a mechanism for ignoring insignificant fea-tures,such surface texture;see Figure 1(b).We have shown that acd can help applications such as skele-tonization [13],perceptually meaningful decomposition [13],and shape deformation [19]to focus on the global shape of the model.Our work in acd has attracted a wide range of interest from the academic community and industry.In particular,we have received many requests to use acd in robot grasping and navigation,Minkoski sum computation,rapid prototyping,and Tele-immersion.Cortical Network Reconstruction .Shape approximation using statistical data can be used to avoid explicit representation of the huge number of neurons and the complexity of each neuron in the cortex [15].We show that the connectivity and geometry of a network that is statistically indistinguishable from an actual cortical network can be efficiently constructed by synthesizing neurons (morphological approximations of real neurons)using L-systems and statistical data and by determining their synaptic connectivity using proximity calculation.As far as we know,this is the first attempt to reconstruct cortical network in this scale.Techniques for Generating and Controlling MotionGenerating and controlling the motion of an object is a fundamental problem in many applications in robotics,animation,and scientific computing.Generating realistic motion with desired characteristics,however,has been known to be extremely difficult.Motion planning provides a tool to generate and control an object’s motion by allowing the user to set initial and final arrangements of the objects and to specify constraints on the motion.Since the motion planning problem has been shown to be intractable,researchers have focused on approximate approaches.The idea behind these strategies is to approximate the topology of the free configuration space (C-space)of a robot by sampling and connecting random configurations to form a graph (or a tree)without explicitly computing the C-space.These approximate strategies have many realistic applications.However,they also have several technical issues limiting their success on some important types of problems,such as the difficulty of finding paths that are required to pass through narrow passages,the difficulty of knowing the required approximation level necessary to solve a problem,and the difficulty of solving the problems efficiently in a large C-space.My work on sampling-based motion planners addresses these issues [18,20,21,23].(uniform)(maprm )Fundamental techniques .We proposed a framework,calledmaprm ,which (approximately)retracts randomly generated configu-rations,free or not,onto the medial axis of the free C-space [18].Weshow improved performance theoretically and experimentally on prob-lems requiring traversal of narrow passages.The figure on the rightshows that maprm can capture the topology of the free C-space whenuniform sampling fails to do so (1000samples in both cases).Our frame-work supports methods with exact or approximate retraction.Exact methods provide fast and accurate retraction in low (e.g.,6)dimen-sional C-space,while approximate methods extend the method to highdimensional problems,such as articulated robots.We also propose an obstacle-based planner [20]to better explore narrow areas.This planner grows a tree to approximate the known free C-space by taking hints from the workspace and C-space obstacles,and is ideally suited to start exploration from maprm samples.Another issue of sampling-based motion planners is that they usually require the user to decide a priori the size of the graph to construct.This decision can be difficult to make,usually requires several attempts,and becomes impractical for time-consuming problems.In recent work,we have addressed this problem by incrementally constructing the graph until some user defined criteria are satisfied[23].Motion of deformable objects .For some problems,the C-space ofan object can be large and complex,thus becoming difficult to approximate.For example,the motion of an object can come from its deformation;see thefigure on the right.Because a deformable object can have many degrees offreedom,generating physically correct motion for deformable objects usuallyresults in intolerable (several hours)running time.We propose the conceptsof hierarchical deformation and an approximate roadmap that allow us toimplement more efficient algorithms that can find perceptually convincingmotion in rigid or deformable environments in just a few minutes [3,19].Techniques for Generating and Simulating Group BehaviorsGroup behaviors are common in nature,such as a school of whales or a colony of ants,and have attracted attention in fields such as robotics,games and animation.Emergent behaviors can generate flocking motion naturally,but more sophisticated group behaviors,e.g.,foraging behavior,cannot be simulated without providing the flock with the ability to memorize,communicate,and reason.On the other hand,giving an agent complete knowledge and information of the system may result in intractable simulations and also unnaturalbehaviors.(coveringbehavior)(shepherding behavior)My key contribution to this area is to integrate traditional locally defined flock-ing behaviors with roadmaps,which provide a global representation of the planningspace,to produce more complex and interesting group behaviors.Our premise isthat the behaviors of an agent are external representations of an agent’s internalknowledge,i.e.,it is the internal representation of the agent that affects its be-havior.We propose the idea of a dynamic roadmap that allows distributed flockmembers to locally modify properties of a roadmap,which is used to represent anagent’s incomplete knowledge of the environment.For example,similar to antsdepositing pheromone,each agent in our system can modify the portion of theroadmap in its vicinity,such as the weight of a nearby edge,to communicate in-directly with other agents.For example,agents can locally update the weight ofthe edges that lead to a certain region of the environment so that other agents canavoid or be attracted to this region.This simple ability allows agents to explorethe environment so that all points of the scene are covered or to locate and bringagents to a goal whose position was not known a priori .These behaviors can be used for mine sweeping and to model foraging behaviorin ants [2].We also demonstrate that,using a roadmap,homing behaviors in whichagents move in a coordinated fashion from one location to another can be specifiedfor the particular surrounding environment [1],e.g.,reducing congestion in narrowpassages.The same technique can also be applied to scenarios involving multiplegroups of agents,such as laser tag,where agents work in teams to score points against agents from opposing teams,and shepherding,where a flock’s directionis steered by outside agents (shepherds);see the figure on the right.These multi-group behaviors have application in many domains,such as games,military strategy training,robots that prevent birds from congregating in airports,cleaning oil spills,and simulating neuron migration.We propose several shepherd locomotions and show that a single shepherd [14]or multiple cooperative shepherds [16]can have better control of the motion of a flock.We are currently developing a framework to ease the process of creating more complex behaviors by automatically composing a set of user defined simple behaviors [17].Please see my research statement for additional details.Our approach has received interest from researchers in robotics [1,14,16,17],animation [4]and artificial life [2,5].Future WorkBoth shape and motion computations play fundamental and critical roles in many fields.In the short term,I would like to focus on applying my approximate convex decomposition concept to problems involving collision detection,shape rendering,shape simplification/compression,and shape identification.I will further investigate efficient motion planning algorithms that cooperate with simplified physical rules to generate natural and convincing motion for complex geometries,such as extremely flexible threads and high degree of freedom characters.The study of these fundamental problems can be used in my long term research on more specific problems in the domains of robotics,computer graphics,computational neuroscience and computational chemistry/biology.Teaching StatementI have always felt enthusiastic about sharing the knowledge that I have learned or discovered with others.I feel excited when the people who I talk to become unpuzzled and even inspired.In fact,this is why I go to school everyday.The other major reason why I want to be a teacher is that,from elementary to graduate school,I was lucky enough to meet several great mentors.They have had profound influences on different perspectives of my study and even life.I am looking forward to this aspect of my academic career and passing on their influences to others.Classroom Teaching.I have been in continuous interaction with undergraduate students since I myself was an undergraduate.As an undergraduate student,I had opportunities to lecture to non-CS students for a Fundamental Computer Concept class and to be a Teaching Assistant for the Data Structure course in my senior year.I have also been a Teaching Assistant for a Pascal class at Texas A&M.My duties as a TA included teaching two labs(4hours/week),designing and grading homework and exams,and holding office hours.Each class had about20students from different majors.I enjoyed being responsible for my labs and made sure that my students did learn.I worked hard to make my teaching inspiring and encouraging.Some of my students became so interested that they even asked for additional work.I also had a student who did poorly at the beginning and asked if he needed a tutor to pass the class.I encouraged him to continue without a tutor and patiently made sure that he understood the materials better and gained confidence.By the end of the semester,he was one of the top students in the class.I felt extremely happy for him.Research Mentoring.Outside the classroom,I have had many opportunities to mentor undergraduates, Masters students and entry-level Ph.D.students in our group and from other universities.I have been the graduate mentor of several undergraduate students participating in research programs over the past three summers.I met with them daily to help them understand what research is,do research,and present their research results formally.Most of the research results have been published as refereed papers[14,16,17]. Ross Sowell,is currently a Ph.D.student at Washington University in St.Louis.As a mentor for other graduate students in our group,I discuss ideas with them,provide suggestions when they are unsure how to proceed with their work,encourage them to work with me,and lead discussions in research project meetings. These experiences have well prepared me for advising my own students as a faculty member.Teaching Philosophy.I believe that the most important thing in teaching is to inspire students.Even if uninspired students work hard and get high scores,they will forget their hard learned knowledge soon after the semester ends.In contrast,inspired students will not only learn in class,they will also learn outside the classroom,from each other,and even after they graduate.I also believe that teaching a student to become an independent thinker is very important.Independent thinking can help students to understand and reason, how to approach a problem,evaluate their approaches,and analyze results instead of memorizing an entire textbook.Even though I had many experiences in teaching and mentoring,I still have a lot to learn to achieve my developed teaching philosophy.I participated in the Graduate Teaching Academy Certification program at Texas A&M to broaden my knowledge/strategies of teaching.I learned how to teach large size classes,teach classes of students with a mixture of backgrounds,and to achieve balance between teaching and research, etc,via a series of seminars provided by several well established teachers from several universities.I also had opportunities to observe different styles of teaching in several courses of Computer Science,Math,and Chemistry,and discussed these teaching styles with my Teaching Mentor.I received a Graduate Teaching Academy fellowship in December2005.Teaching Interests.I have a wide range of teaching interests.I can teach any undergraduate class, any core graduate class and specialty graduate courses in my areas of expertise.I look forward to giving back the knowledge I acquired through the years of my study,especially the topics related to computational geometry,computer graphics and animation,algorithms,and robotics.References[1]O.B.Bayazit,J.-M.Lien,and N.M.Amato.Betterflocking behaviors using rule-based roadmaps.InProc.Int.Workshop on Algorithmic Foundations of Robotics(WAFR),pages95–111,Dec2002. [2]O.B.Bayazit,J.-M.Lien,and N.M.Amato.Better group behaviors in complex environments usingglobal roadmaps.In Artif.Life,pages362–370,Dec2002.[3]O.B.Bayazit,J.-M.Lien,and N.M.Amato.Probabilistic roadmap motion planning for deformableobjects.In Proc.IEEE Int.Conf.Robot.Autom.(ICRA),pages2126–2133,May2002.[4]O.B.Bayazit,J.-M.Lien,and N.M.Amato.Roadmap-basedflocking for complex environments.InProc.Pacific Graphics,pages104–113,Oct2002.[5]O.B.Bayazit,J.-M.Lien,and N.M.Amato.Swarming behavior using probabilistic roadmap techniques.Lecture Notes in Computer Science,2005(3342):112–125,January2005.[6]T.-Y.Li,J.-M.Lien,S.-Y.Chiu,and T.-H.Yu.Automatically generating virtual guided tours.In Proc.IEEE Computer Animation(CA),pages99–106,May1999.[7]J.-M.Lien and N.M.Amato.Approximate convex decomposition for putational Geom-etry:Theory&Applications.To appear.[8]J.-M.Lien and N.M.Amato.Approximate convex decomposition.Technical Report TR03-001,ParasolLab,Dept.of Computer Science,Texas A&M University,Jan2003.[9]J.-M.Lien and N.M.Amato.Approximate convex decomposition.In Proc.20th Annual ACM Symp.Computat.Geom.(SoCG),pages457–458,June2004.Video Abstract.[10]J.-M.Lien and N.M.Amato.Approximate convex decomposition of polygons.In Proc.20th AnnualACM putat.Geom.(SoCG),pages17–26,June2004.[11]J.-M.Lien and N.M.Amato.Approximate convex decomposition of polyhedra.Technical ReportTR05-001,Parasol Lab,Dept.of Computer Science,Texas A&M University,Jan2005.[12]J.-M.Lien and N.M.Amato.Polyhedron realization using convex hull projection.Technical ReportTR05-016,Parasol Lab,Dept.of Computer Science,Texas A&M University,Dec2005.Submitted for publication.[13]J.-M.Lien and N.M.Amato.Simultaneous shape decomposition and skeletonization.Technical ReportTR05-015,Parasol Lab,Dept.of Computer Science,Texas A&M University,Dec2005.Submitted for publication.[14]J.-M.Lien,O.B.Bayazit,R.-T.Sowell,S.Rodriguez,and N.M.Amato.Shepherding behaviors.InProc.IEEE Int.Conf.Robot.Autom.(ICRA),pages4159–4164,April2004.[15]J.-M.Lien,M.Morales,and N.M.Amato.Neuron prm:A framework for constructing cortical networks.Neurocomputing,52-54(28):191–197,June2003.[16]J.-M.Lien,S.Rodriguez,J.-P.Malric,and N.M.Amato.Shepherding behaviors with multiple shep-herds.In Proc.IEEE Int.Conf.Robot.Autom.(ICRA),pages3413–3418,April2005.[17]J.-M.Lien,S.Rodriguez,X.Tang,J.Maffei,D.Corlette,A.Masciotra,and posablegroup behaviors.Technical Report TR05-006,Parasol Lab,Dept.of Computer Science,Texas A&M University,Sep2005.[18]J.-M.Lien,S.L.Thomas,and N.M.Amato.A general framework for sampling on the medial axis ofthe free space.In Proc.IEEE Int.Conf.Robot.Autom.(ICRA),pages4439–4444,September2003.[19]S.Rodriguez,J.-M.Lien,and N.M.Amato.Planning motion in completely deformable environments.Technical Report TR05-010,Parasol Lab,Dept.of Computer Science,Texas A&M University,Sep2005.To appear in Proceedings of the IEEE International Conference on Robotics and Automation(ICRA), Orlando,May2006.[20]S.Rodriguez,X.Tang,J.-M.Lien,and N.M.Amato.An obstacle-based rapidly-exploring randomtree.Technical Report TR05-009,Parasol Lab,Dept.of Computer Science,Texas A&M University,Sep 2005.To appear in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA),Orlando,May2006.[21]A.Vargas,J.-M.Lien,M.A.M.A.,S.Rodriguez,and er-guided path planning.Technical Report TR05-011,Parasol Lab,Dept.of Computer Science,Texas A&M University,Sep 2005.[22]A.Vargas,J.-M.Lien,and N.M.Amato.vizmo++:a visualization,authoring,and educational toolfor motion planning.Technical Report TR05-014,Parasol Lab,Dept.of Computer Science,Texas A&M University,Sep2005.To appear in Proceedings of the IEEE International Conference on Robotics and Automation(ICRA),Orlando,May2006.[23]D.Xie,S.L.Thomas,J.-M.Lien,and N.M.Amato.Incremental map generation(IMG).TechnicalReport TR05-007,Parasol Lab,Dept.of Computer Science,Texas A&M University,Sep2005.。