《VTARC：2024具身智能未來方向研討會報告（英文版）（33頁）.pdf》由會員分享，可在線閱讀，更多相關《VTARC：2024具身智能未來方向研討會報告（英文版）（33頁）.pdf（33頁珍藏版）》請在三個皮匠報告上搜索。

1、Future Directions Workshop on Embodied IntelligenceMay 1920,2024Joshua Bongard,University of VermontKyujin Cho,Seoul National UniversityYong-Lae Park,Seoul National UniversityRobert Shepherd,Cornell UniversityPrepared by:Kate Klemic,Virginia Tech Applied Research CorporationSithira Ratnayaka,Virgini

2、a Tech Applied Research CorporationFuture Directions Workshop seriesWorkshop sponsored by the Basic Research Office,Office of the Under Secretary of Defense for Research&Engineering CLEARED For Open Publication Department of Defense OFFICE OF PREPUBLICATION AND SECURITY REVIEW Oct 09,2024ii Contents

3、Preface iiiExecutive Summary 1Introduction 5Research Challenges 7Research Opportunities 12Conclusion 17Bibliography 18Appendix I Workshop Attendees 19Appendix II Workshop Agenda and Prospectus 27iiiInnovation is the key to the future,but basic research is the key to future innovation.Jerome Isaac Fr

4、iedman,Nobel Prize Recipient(1990)PrefaceOver the past century,science and technology has brought remarkable new capabilities to all sectors of the economy;from telecommunications,energy,and electronics to medicine,transportation and defense.Technologies that were fantasy decades ago,such as the int

5、ernet and mobile devices,now inform the way we live,work,and interact with our environment.Key to this technological progress is the capacity of the global basic research community to create new knowledge and to develop new insights in science,technology,and engineering.Understanding the trajectorie

6、s of this fundamental research,within the context of global challenges,empowers stakeholders to identify and seize potential opportunities.The Future Directions Workshop series,sponsored by the Basic Research Office of the Office of the Under Secretary of Defense for Research and Engineering(OUSD(R&

7、E),seeks to examine emerging research and engineering areas that are most likely to transform future technology capabilities.These workshops gather distinguished academic researchers from around the globe to engage in an interactive dialogue about the promises and challenges of each emerging basic r

8、esearch area and how they could impact future capabilities.Chaired by leaders in the field,these workshops encourage unfettered considerations of the prospects of fundamental science areas from the most talented minds in the research community.Reports from the Future Direction Workshop series captur

9、e these discussions and therefore play a vital role in the discussion of basic research priorities.In each report,participants are challenged to address the following important questions:How will the research impact science and technologycapabilities of the future?What is the trajectory of scientifi

10、c achievement over the nextfew decades?What are the most fundamental challenges to progress?This report is the product of a workshop held May 19-20,2024 in Seoul,South Korea on the future of Embodied Intelligence research,as an essential and critical aspect of future robotics that are agile and endu

11、ring,as well as damage tolerant.It is intended as a resource to the S&T community including the broader federal funding community,federal laboratories,domestic industrial base,and academia.1 Executive SummaryEmbodied Intelligence(EI)is a rapidly evolving field that seeks to address new ideas about t

12、he nature of machine intelligence.EI blurs the lines between Artificial and Physical Intelligence(AI and PI,respectively);it creates a diffuse interface between artificial and natural components of a system.EI aims to incorporate into machines the multimodal and multiscale adaptation observed in nat

13、ural organisms,for a wholly new approach to robotic technology,allowing a future filled with autonomous,useful,and safe machines.Consider a world in which every machine is morphologically and neurologically unique.Such technologies would be immune to unintentional surprise(novel environments)or inte

14、ntional surprise(adversarial attacks)because no two machines would share a common Achilles heel.Imagine machines that,when cleaved in two,form two smaller yet distinct versions of the original machine.Imagine machines that can devolve into swarms of independent components and reform into a physical

15、unity on demand.Consider autonomous machines in which there is no clear distinction between control,actuation,sensation,communication,computation,and power,rendering such machines immune to complete failure of any one sub-system.These machines might also incorporate living and non-living components,

16、further combining the best of the biotic and abiotic worlds and blurring the distinction between“us”(humans)and“them”(machines).Past efforts in embodied intelligence science have proceeded with little interaction between the broad fields in which R&D was pursued.In the present,there are many efforts

17、 surrounding the integration of new AI and machines,leading to a need to integrate the brain and body of these systems.Biological systems have served as inspiration for many of the modern applications and for these systems.There is an opportunity within embodied intelligence to cause responses inter

18、mediate to pre-flex,reflex,and centralized decision making.An octopus tentacle provides a good example of a higher level,yet localized Observe,Orient,Decide,Act(OODA)loop.While the whole octopus has a centralized brain,it also has a large number of neurons in its tentacles.Even when separated from t

19、he body,the suckers of a tentacle are able to sense the chemical environment,locally decide if an object is food or not,grasp it if it is food and pull it towards its concept of where its beak is located.Sensing,actuation,computation,and energy is distributed through every mm3 of an octopuss flesh;a

20、 living,autonomous material system.While AI excels at handling large amounts of data,the underlying statistical process of learning is not conducive to causal and abstract reasoning.Attempts to create such capability within that framework have generally not yielded consistently accurate results,and

21、this likely relates to the difference between how engineered(AI)and natural organisms learn.These are fundamental questions:what is learning and,to a deeper level,what is intelligence?Advances in these fields could lead to further integration between humans and machines,creating new ecosystems in wh

22、ich all can co-exist.The Future Directions Workshop on Embodied Intelligence was held on 19-20 May 2024 in Seoul,South Korea to examine the prospects for applying new approaches,theories,and tools in basic research to enable these capabilities over the next 10-20 years.It gathered 28 researchers fro

23、m a variety of fields,including soft robotics,motion control,biomechanics,mechanical engineering,control theory,systems biology,physics,mathematics,computer science,and bioethics.The workshop included researchers from the Republic of Korea(ROK)and the United States(US)and served as a foundation for

24、collaboration in the field between the two countries.The workshop was organized for highly interactive small group discussions with whole-group synthesis on the challenges,opportunities and trajectory of research across three pillars of embodied intelligence:perception,motion,and adaptation.Research

25、 Challenges Participants identified the following challenges for each technical pillar of EI and identified the key research areas to realize the envisioned future of embodied intelligence.PerceptionThe ability for machines to sense their surroundings/environment and glean information from it has be

26、en explored utilizing several methods.Several sensing modalities have emerged that interface the body with the environment(exteroception)and provide more detailed knowledge about the bodys state internally(proprioception).Perception goes beyond this by incorporating senses such as olfactory(smell,or

27、 the ability to detect chemical information)and nociception(the ability to detect harmful environmental stimuli),which are more exotic methodologies that can be utilized for environmental navigation.The main research challenges for Perception are linked to the fundamental questions surrounding sensi

28、ng and the ability to infer information from sensors.For all natural organisms,knowledge representation is highly dependent on the sensory modes,and their processing and fusion.Thus,learning cannot be dissociated from the sensors used to acquire information.In addition,using sensing in the artificia

29、l world revolves around vision,while in the natural world,a plethora of other methods are used.The main Perception research challenges include:Sensitivity:Increasing the signal to noise ratio by localizingsignals of interest and amplifying is a dynamic and computationally challenging process that ha

30、s the potential to increase agility and energy efficiency if performed at the embodied level.2Innervation:Multiplexing many sensors and laying themout sensibly inside a complex structure is a manufacturing challenge that has the potential to greatly increase perceptive capabilities with the ability

31、to localize all sensing and actuation.Encoding:Achieving high information throughput by siftingthrough large data amounts effectively by leveraging optical modes,biological spiking,etc.is a data challenge that has the potential to provide massive data rates for information fusion without exploding t

32、he practical wiring and assembly requirements to sensing hardware.MotionTraversing and navigating the environment is a staple of any system/body.This feat is done directly by locomotion,or by changing the environment to suit the needs of the system/body.The degrees of freedom(DOF)are directly correl

33、ated to the complexity of the system,but can change over its lifetime,providing increased adaptability for movement.The main research challenges for Motion are linked toward the current technological trends to dominate the environment rather than leveraging it.Being able to utilize it effectively wi

34、ll greatly increase the energy efficiency and synergy of the body and the environment.The main motion research challenges are listed below:Agility:Increasing responsiveness and power,withoutincreasing DOF,will need to be done by supplying power and data to actuators;mimicking natures bottom-up appro

35、ach of self-assembly allows for far more architectural complexity.Endurance:Withstanding many cycles of use,or usingless energy for operations by being efficient,will need the utilization of multifunctional energy storage and transduction,high energy density fuels,storage and release of elastic ener

36、gy,and center of mass adjustments during locomotion.Growth:Changing to the environment(or changing theenvironment)by adding,subtracting,or changing dimensions,body segments,and/or DOF with increased ability to utilize energy will be a major challenge for the machines body.AdaptationThe natural world

37、 has solved many design problems via evolution.Artificial systems can be imbued with this capability by utilizing a wide variety of computational techniques designed to optimally modify the body to the environment.An additional feature to be explored is the use of collective adaptation,in which many

38、 bodies act as a whole to perform specific tasks,and thus can be changed to better fit their environment.The main research challenges for Adaptation are centered around the co-design of brain and the body.While natural systems use evolution,artificial ones adapt from centralized computation.Efficien

39、t management of energy expenditure also will be challenging,but taking advantage of materials science and additive manufacturing may ameliorate these engineering contradictions.The main adaptation research challenges are listed below:Learning:Logic links will need to be increased basedupon new exper

40、iences and will extend beyond traditional neural plasticity to the bodies of robots.The bodies as well as the brains of future robots may learn how best to detect co-occurring features of external challenges(or internal challenges,such as injury),and prepare themselves morphologically and neurologic

41、ally to handle those challenges when the re-occur.Language:Verbal claims could be demonstrated physicallyas a self-correcting mechanism for confabulation;this task could use Large Language Models(LLMs)as a supplement,but not a sole use as they suffer from hallucinations and the generation of non-fac

42、tual verbal statements.Control:By adding DOF(and reducing the discreetboundaries between the body and the environment),the control of the systems will be a challenge;there will be kinematic redundancy for systems with too many DOF for their tasks(but adding the appropriate DOF will allow for more fl

43、exibility);this inefficiency will need to be addressed by selectively removing DOF(or adding more).A Tapestry of ChallengesPerception,Motion,and Adaptation are interdependent topics that will require concurrent research efforts.Subjects such as information density will need to be addressed utilizing

44、 all three to be effective:Perception to amplify or filter data,Adaptation to understand the resulting information,and Motion to adjust for it.Indeed,organisms change based on their environments utilizing all three of these.In biology,organisms focus on relevant stimuli utilizing sensing organs and

45、develop behavioral responses which filter out unimportant inputs.They respond based on the organisms needs,based upon its internal state and due to limited attention/energy.They leverage the past experiences of the organism via learning and memory,which leads to innate responses and reflexes which h

46、elp save energy and assist in remodeling and growth of the organism.The interdependent nature of artificial systems also requires feature integration.The system first selects the features that it deems useful,then extracts them(or their information).In order to use the new features,the system will t

47、hen regularize the newly acquired features and optimize algorithms to reproduce the features for use by the system.Unlike biological systems,however,these processes consume large amounts of energy and are subjected to significant latency.To tackle these interconnected research challenges,a concerted

48、 effort must be made to foster collaboration and communication among researchers in diverse fields.Increasing knowledge transfer between groups of researchers with defined taxonomy and common language is a first step to this goal.Concerted 3 Testing,Evaluation,and Validation(TEV)will also be paramou

49、nt to realizing this objective.Transdisciplinary research,which includes materials science,manufacturing,computer science,mechanical engineering,and EI design will need to be woven together.Research OpportunitiesAs engineering advances produce ever more sophisticated artificial systems,there are tre

50、mendous research opportunities to learn from biological ones.Indeed,organismal biology already shows the ability to focus on relevant stimuli,respond based on needs,and leverage past experiences.These systems can also be studied to observe their ability to identify and integrate new features from th

51、e environment,perhaps revealing key insights to be able to translate such features to synthetic systems.With the advancements of other fields,there exists many opportunities for exciting developments and research to be conducted in the field of embodied intelligence.Some of these include additive ma

52、nufacturing,neuromorphic computing,biohybrid robotics,autonomous material systems,and electrochemistry.Research TrajectoryThe workshop participants developed a trajectory for the research opportunities identified for the field of embodied intelligence with a vision for the 5-,10-,and 20-year horizon

53、s.Five-year goals In the immediate future,EI will augment existing robot architectures.These robots,equipped with simple control loops informed by analog sensing and processing layers,commanding actuators(e.g.,continuum,compliant,standard)will be capable of reduced energy expenditure during mobility

54、 tasks or more dextrous performance in assembly tasks for example.These robots may feature,as an example,endoskeletal structures with soft actuators and skins,mediating reconfigurability based on task requirements.The use of compliant manipulators and soft skins will improve their agility and endura

55、nce compared to non-EI systems.Key Goals:Develop consensus metrics for energy consumption duringstate transitions(e.g.,trotting to cantoring),as well as agility(e.g.,acceleration and turn radius).Establish foundational control strategies using logical basisfunctions for coordination of tasks.Over th

56、e next decade,EI is expected to leverage prior results in analog sense-act-respond functions to produce a set of low level robots that demonstrate these principles with specific functions,akin to organs or“polyps”seen in biology(Figure 5).The results may be akin to reconfigurable systems of modules

57、mediated by analog computational layers that can configure for(as an example)external dexterity or(another example)internal operational efficiency for existing tasks.Importantly,the development of basis functions for the set of modules will play a critical role in this phase,allowing robots to be dy

58、namically assembled and disassembled in response to environmental or task changes.Key Goals:Enumeration of agility and endurance requirements forgeneral purpose robotics(these numbers should be arrivedat beyond just EI community)Define a set of low-level EI modules that address therequirements for a

59、gility and endurance Algorithms developed that provide the basis forcoordination between these modules(digital and analogsolutions)Long-Term(20 Years)In the long term,EI researchers will understand how to best leverage living and synthetic approaches to build low-level EI modules.The basis functions

60、 to coordinate the low-level biohybrid robots to autonomously assemble and disassemble themselves into more complex,high-level robots will be known.These high-level robots are more sophisticated,capable of performing complex tasks and adapting to changing environments.This synthesis will enable the

61、development of general-purpose robots capable of growth,reconfiguration,and continuous adaptation.Logical basis functions(e.g.,autonomous material computation)(Yamada et al.,2022)will be fully integrated into the robots architecture,enabling seamless coordination across multiple robots in various en

62、vironments.In addition to the coordination of low-level robots,we also anticipate autonomous coordination between multiple(and different)high-level robots.Key Goals:Develop autonomous material systems(AMS)that allow forindependent sensing and dynamic reconfiguration.Implement neuron-based computing

63、for acceleratedadaptation and coordination of large robot assemblies.Advance multiplexed high-DOF actuator arrays to supportsophisticated motion and structural integrity duringassembly and disassembly.Robust approaches to maintaining life in real worldenvironments,as well as mediating their interfac

64、e withartifices.Communication protocols in addition to RF and visualspectrum signaling,such as acoustic and chemical.Opportunities to Achieve these GoalsThis workshop report outlines the opportunities and a path forward for research in the field of embodied intelligence.One aspect is the utilization

65、 of DOF,both to manipulate and understand the limitations,that will be integral to the advancement of the field.Challenges of manufacturing and computational efficiency must be addressed alongside long-term testing protocols and energy considerations.A concerted effort must be made to bring together

66、 the community to address these challenges through interdisciplinary research and collaboration.Improving communication and idea-sharing within the community is imperative for the future of this field.The participants emphasized the importance of the following technology areas:4Materials and Manufac

67、turing:Advances in materials andmanufacturing will enable robots to be designed with more heterogeneous materials which reduce and eventually eliminate the need for subsystems.Autonomous Material Systems will allow for a high degree of adaptability in robots,with reduced cost in manufacturing.Adapta

68、tion and Computation:Advances in computationalhardware will enable hyper-efficient computation systems that integrates seamlessly with physical substrates,enabling more efficient and adaptive behaviors.These systems will operate beyond current digital communications and memory and rather use analog

69、and biotic computation with enhanced response speeds and/or reduced power consumption.As more complex Large Language Models(LLMs)are developed and integrated with other interaction modes,the communication between Perception,Adaptation,and Motion domains will become more efficient and capable,allowin

70、g for higher complexity and compute.Application Focus Areas There exist many applications for these new technologies.Focus areas for these use-cases include:Daily Life and Labor Replacement:Society will thrive in a newera in which robotic assistance reduces human work,addressing labor shortages and

71、removing risk from human workers.Healthcare and Robotics:Affordable soft robots for patient carewill allow for precise and enhanced patient recovery,with hard exoskeletons utilized for rehabilitation and emergency response.Advanced Task-Specific Robots:Low-cost robots will beavailable for unique tas

72、ks,which will be simpler but more effective than current robots.Accelerating the FieldThe participants discussed means for accelerating the field.They note that an increased focus on partnerships with industry will yield more efficient and viable advances.Key enablers include:Collaboration and Commu

73、nity:Interdisciplinary collaborationbetween robotics,biology,AI/ML will need to develop to lay the foundation for ubiquitous use of robots in society.Training programs will also need to mirror these collaborations,with holistic and comprehensive learning and teaching of the next generation of resear

74、chers.Metrics and Evaluation:standardized testing and assessmentwill be necessary to streamline the advancements in the field.A DARPA Robotics Challenge for Embodied Intelligence,for example,would push the frontiers of robotics by promoting integration of Embodied Intelligence within existing robots

75、.Successful projects that displayed true mastery of perception,motion,and adaptation with low energy expenditures would be crucial to drive forth future Embodied Intelligence research and development.5 IntroductionThe idea that the body and brain are separated is an assumption about machine intellig

76、ence that was formed in the distant past but continues to constrain how we approach AI and robot technology development today.This report is an attempt to formulate a new view of embodied intelligence,free of prior assumptions,to promote step changes in robotics.We anticipate progress in this domain

77、 will dramatically improve agility,endurance,and damage tolerance in our automated machinery.We note that in this emerging field,the terminology used to label it is confusing:Artificial Intelligence,Physical Intelligence,Embodied Intelligence,as well as several other phrases are used synonymously an

78、d sometimes antonymously.To aid in reading this report,we make a brief attempt at clarifying some of the more important terms:Artificial Intelligence(AI)is used to describe algorithmsrepresented,typically,in software that provides output(e.g.,recommendations or actuations)based on inputs(e.g.,instru

79、mented measurements or human suggestions).Physical Intelligence(PI),is used for robots that haveAI embedded in firmware and operating locally onautonomous hardware as well as a synonym for EmbodiedIntelligence.For the former definition,PI tends to assumea thermodynamically closed machine(the mass an

80、d energyavailable to the machine come from within).Embodied Intelligence(EI)is used to describe systemsthat blur the lines between the machines body and theenvironment in which it is interacting;ultimately,it will be ananalog approach to interacting with the world.EI envisionsthermodynamically open

81、machines that can incorporate newmass and energy to recover or expand their capability.Embodied Intelligence blurs the interface between machine and environment,and between the boundaries of internal components or modules.In external interactions with the environment,for example,EI systems allow the

82、 gravel below a robots foot to change the shape of the footstoring energy,adding stability,and becoming part of the machine for milliseconds prior to release.A projectile impacting a surface may partially and reversibly imbed itself into the volume,the newly formed object can make a decision whether

83、 to accept or reject the new form at the speed of sound.A chemical spray may change the macromolecular orientation of the surface,changing its optical and mechanical properties,displaying a warning to human teams and changing the trajectory of a robot away or towards the source.Internal to a machine

84、,multiple interacting low-level robotic subsystems could synthesize more complex autonomy;this function is seen biologically in zootic animals such as the Portuguese man owar.The past.EI is a broad but as yet disjointed effort to heal a millennia-old assumption in Western thought,which is that the m

85、ind and body are distinct.The late Daniel Dennett referred to this bias in Western thinking as Cartesian gravity:it is so ubiquitous that it usually escapes notice,yet it influences the action of everyone and everything.Researchers attempting to create intelligent technologies are not immune to this

86、 pull.Although a philosophical bias dating back to Descartes and Plato,such“Cartesian Dualism”influences how the research community currently approaches the creation of autonomous and safe machines.Proof of Cartesian Dualisms influence can be seen in the bicameral shape of the field itself.Researche

87、rs tend to work on purely non-physical“AI”technologies such as large language models,or physical machines such as robots or autonomous vehicles.In practice,there is little overlap between researchers in these fields.Eastern thought tends to adopt a more holistic stance to the natural world,and to in

88、telligent organisms by extension:no obvious distinction is made between the mechanical,chemical,and electrical supports of intelligent behavior in humans or animals.This fact alone demands a better integration between Western and Eastern researchers in the basic approach to realizing intelligent mac

89、hines that use their bodies,as well as their control policies to realize useful and safe behavior.The present.Evidence increasingly demonstrates that non-embodied and embodied approaches to autonomous and safe machines need each other.Software LLMs are now capable of facilitating a wide range of use

90、 cases,but they are dangerous:no guarantees exist that they will not err or fabulate in a way that escapes the human users notice,especially in applications where human safety is involved directly(i.e.seeking medical advice from a chatbot)or indirectly(AI-generated code that controls medical equipme

91、nt).Conversely,autonomous robots are increasingly reliable,but only within very narrow applications and environments,such as autonomous driving on pedestrian-free roads in normal lighting and dust-free air.Living systems,in contrast,are capable of handling internal surprise(injury)and external surpr

92、ise(novel stimuli)while performing a wide range of tasks such as feeding,migrating,or problem solving,in a wide range of environments.Organisms balance generality and safety by generating behavior as a function of their bodies and nervous systems at a deep level.The ways in which they achieve this a

93、re only now becoming clear.Channeling such discoveries from nature into machines could pave the way toward a future populated by complex,general-purpose and capable machines that can work safely alongside,and even inside,humans,but only if this integration is done correctly.There is a growing intere

94、st in Physical Intelligence,or“Embodied AI”,in basic research labs and applied technology companies.Usually,in such cases,no regard is given to how natural systems deeply integrate electrical,chemical,and mechanical adaptation at all spatial and temporal scales.Instead,non-physical foundation models

95、 are dropped into machine“shells”that have a few components capable of adaptation,such as motors and sensors,but are otherwise built from inert materials such as metal and plastic.Such superficial couplings could lead to the worst 6of both worlds rather than the best of both worlds:robots could inhe

96、rit foundation models unpredictability when confronted with novel stimuli,and robots with fixed bodies that generate narrow sensorimotor experiences could narrow the understanding of non-embodied AIs trained on that data.Thus,there is a pressing need for basic research to understand“how”best to inte

97、grate mind and body in machines.The future.If discoveries about how organisms realize multimodal and multiscale adaptation could be successfully incorporated into machines,a wholly new future filled with autonomous,useful,and safe technologies becomes possible.Consider a world in which every machine

98、 is morphologically and neurologically unique.Such technologies would be immune to unintentional surprise(novel environments)or intentional surprise(adversarial attacks)because no two machines would share a common Achilles heel.Imagine machines that,when cleaved in two,form two smaller versions of t

99、he original machine.Imagine machines that can devolve into swarms of independent components and reform into a physical unity on demand.A common recombinant basis for their functional synthesis would be defined.Consider autonomous machines in which there is no clear distinction between control,actuat

100、ion,sensation,communication,computation,and power,rendering such machines immune to complete failure of any one sub-system.Consider autonomous vehicles that effortlessly switch between visual navigation in normal conditions,inertial navigation in dust-choked air,thermotactic navigation in smoke-fill

101、ed air,chemotaxis for chemical spill escape,and biological sensing in pathogen-laced air.Consider machines that are unique combinations of living and non-living components,further combining the best of the organic and inorganic worlds and blurring the distinction between“us”(humans)and“them”(machine

102、s).Such technologies would not become an additional layer of unpredictable actors on top of an already complicated society.They would become reliable due to interdependence;they would become a reliable ecosystem among themselves,and with the natural and human worlds.The above is not science fiction,

103、but extrapolations from current theory and physical prototypes.This vision is what Embodied Intelligence could be,and how it could support and enrich society,if basic research thoroughly investigates the intertwined roles of physicality and cogitation in nature,and how best to translate that unity i

104、nto machines.7 Research ChallengesEmbodied Intelligence expands the computational framework of biological and artificial autonomy beyond a centralized computer(e.g.,brain or microchip)and into the architecture of the body.This embodiment of intelligence is aided by a theory of“morphological computat

105、ion,”where the input of environmental stress is processed by the materials and structures,reducing computational load on a central computer for a command response,or negating the need for a traditional computing architecture altogether(i.e.,reflex action).The last decade has seen proliferation of th

106、ese concepts,publications and citations seeing exponential growth.In parallel,new advances in material science,neural networks,soft robotics,biohybrids,additive manufacturing,parallel computing,and signal processing have made it important to revisit and refine these concepts.With these new scientifi

107、c models and technologies come new opportunities and challenges.To contextualize the challenges,we have created three technical pillars with exemplar sub-domains that give specific examples of areas in which research must be performed to realize the future we have envisioned.We briefly define what w

108、e intend by the sub-domain names in Figure 1:Figure 1.Hierarchy of topics related to brain/body co-evolutionPerceptionInnervation Distributing sensing,communication,andcomputation into the volume of the machineEncoding The process of converting information into aformat that can be stored,transmitted

109、,or processed by a robotSensitivity The ability of a robot to detect and respond tosubtle changes or stimuli in its environmentMotionGrowth Adding or subtracting mass in response toenvironmental triggers or timeAgility Moving quickly and nimbly,often in response tochanging conditions or unexpected o

110、bstaclesEndurance The ability of a robot to operate continuously forextended periods without needing maintenance or rechargingAdaptationControl The ability to command and regulate the behavior ofa robotLanguage Contextualizing interactions between humans andother robotsLearning Acquiring new knowled

111、ge or skills throughexperience or interaction with its environmentPillar 1:PerceptionWhile AI excels at handling large amounts of data,the underlying statistical process of learning is not conducive to causal and abstract reasoning.Attempts to create such capability within that framework have genera

112、lly not yielded consistently accurate results,and this likely relates to the difference between how engineered(AI)and natural organisms learn.These are fundamental questions:what is learning,and to a deeper level,what is intelligence?For all natural organisms,knowledge representation is highly depen

113、dent on the sensory modes,and their processing and fusion.Thus,learning cannot be dissociated from the sensors used to acquire information.A variety of sensing modalities have emerged that interface the body with the environment(exteroception)and provide more detailed knowledge about the bodys state

114、 internally(proprioception).These sensors are typically fused with traditional analog-to-digital converters for processing by standard computer architectures,but there is an opportunity within embodied intelligence to cause responses more like reflex actions using analog computation,and using a vari

115、ety of fields(electrical,magnetic,mechanical,chemical,etc.).Sometimes these sensors may also be fused with computation,preprocessing information;a human eye,for example,not only measures the properties of light,but it also performs preprocessing functions akin to wavelet transforms.Artificially,we u

116、se vision almost exclusively for navigation.Nature,however,is not so reliant on vision.There are many examples of complex organisms that do not use vision(Figure 2);however,there are no examples of animals we are aware of that do not use touch.A blind mole rat navigates intricate tunnel systems,a Ka

117、uai cave spider actively hunts by feeling its preys vibration signature,and blind Mexican Tetras can school by feeling the complex hydrodynamic interactions of the group.This huge discrepancy in perception between the artificial and natural world is an example of how EI will leverage unused environm

118、ental cues for improved maneuverability and efficiency.These natural examples of navigation,hunting,and schooling by feel could create the basis functions for the constitution of independent low-level robotic systems into larger,high-level physical agents.Growth Agility Endurance Control Language Le

119、arning Innervation Encoding SensitivityMotionAdaptation PerceptionBrain/BodyCo-Evolution8Figure 2.Examples of complex organisms that do not use vision.Source:Wikipedia,2024Within this pillar of EI,we have identified non-exhaustive grand challenges for research.Sensitivity is certainly one of these c

120、hallenges,typicallymeasured in Signal to Noise Ratio with units of decibels.Localizing signals of interest and amplifying them above those not important at the time is a dynamic process and computationally challenging;performing it at the embodied level would increase agility and energy efficiency.I

121、nnervation of machine volumes to assess the environment andinternal state of the machine is a manufacturing challenge.As the most obvious way to increase perceptive capabilities is with more sensors,simple wiring of them will become intractable without new materials and manufacturing methods.Autonom

122、ous Material Systems have the potential to localize all sensing and actuation.Encoding data for high informationthroughput(e.g.,bits/s;bits/s/W;bits/s/cm3;bits/s/kg)is another challenge.Leveraging optical modes,biological spiking,etc.would be an important way to provide massive data rates for inform

123、ation fusion without exploding the practical wiring and assembly requirements to sensing hardware.Pillar 2:Motion Intelligence is a developmental process:(i)within an organisms lifetime and(ii)throughout a species evolution.In the first example,the“curse ofdimensionality”is somehow solved bynature,w

124、hereby complex organisms learn to control their large number of degrees of freedom(DOFs)using large numbers of sensory inputs.An interesting hypothesis as to how nature accomplishes this task is that through the developmental process,DOFs are initially frozen and released,or added with growth.Furthe

125、rmore,the organism is not passive;it can actively probe and modify the environment,using various actuators and tools.Learning and,by association,intelligence,are a function of available modes of sensing and action.For example,human babies learn their environment sequentially,according to their abili

126、ty to move and manipulate.This creates a challenging co-design problem for robotics,for which mechanical operation,sensors,neural processors and training/learning strategies must then be designed concurrently,and these concurrent designs derivative of prior instantiations.The components will grow an

127、d rearrange,grow and shrink,over time.Examples of how research in motion can reduce the disparity between our present artifices and nature via EI are shown in Figure 3.Figure 3A shows how we presently use our technology to dominate the environment(ailerons to manage turbulence)whereas birds leverage

128、 turbulence to save energy.Figure 3B shows how we engineer our landscape to be flat to handle tires,but a frog will store and release elastic energy in the environment(e.g.,leaves)to save energy and traverse complex terrain.Agility is not only a hard-to-define characteristic;it is a featthat is diff

129、icult to achieve artificially.Increasing the acceleration and deceleration of objects with precision trajectories requires Kauai Cave Wolf SpiderLesser Blind Mole-RatMexican Tetra(Blind Cave Fish)School of Mexican TetraFigure 3.Examples of research in motion to improve the relation between artificia

130、l and naturalsources.A:Examples of how to use our technology or to dominate the aerial environment in contrast to birds.B:Examples of how we engineer our landscape to be flat to handle tires in contrast to a frog which stores and releases elastic energy in the environment.Sources:Laurent et al.2021,

131、Kirstines.Dk.(2016,June 14)9 high power actuators and high number of DOFs,while being lightweight.Supplying power and data to these actuators is a difficult manufacturing challenge from the top down.Natures bottom-up approach of self-assembly allows for far more architectural complexity;e.g.,the co-

132、existence of neurons for sensing and information processing along with muscle for actuation in every mm3 of an octopuss tentacle.Endurance while being agile is a paradox ripe for improve withEI.Animals are far more capable than our machinery of being highly responsive to the environment while being

133、able to operate for days and weeks without additional fuel consumption.The challenges here revolve around multifunctional use of energy storage and transduction,high energy density fuels,storage and release of elastic energy,and center of mass adjustments during locomotion(Aubin et al.,2022).Growth

134、of the machines body to allow for learning,circumventing obstacles,or manipulating objects is a challenge.The challenge is to build a machine that can change dimensions,add or remove body segments,freeze or add DOFs,or even literally grow are material science and manufacturing challenges.Pillar 3:Ad

135、aptation In nature,this co-design problem is solved through evolutionary processes.Indeed,hardware also evolves over time to account for inefficiencies in design or changing end-user needs.Using sim2real(the process of transferring skills learned in simulation to real-world applications),evolutionar

136、y algorithms,and other advanced computation-based techniques,we can better design autonomous systems that are more adaptable to changing environments,perhaps an organisms best indicator of intelligence.This ability to tune the energy landscape of the autonomous system,and impedance match it to envir

137、onmental inputs and outputs,is at the core of embodied intelligence.Indeed,another true measure of intelligence beyond mammals and bird examples would not be the capability of expending huge amounts of energy,but managing it instead.By taking advantage of materials science,additive manufacturing,or

138、building new approaches,this artificial species ability to tune the I/O and energy landscape can be evolved more rapidly.Adaptation,artificially,is primarily achieved from centralized computation(Figure 4).However,biology relies far more on lower order feedback loops and structural organization.From

139、 examples like the peripheral nervous system,to colonial organisms,to even collective work from swarms for simpler organisms.The Portuguese man owar,an example of a colonial organism(i.e.,zooid)comprised of different species,have separate chemical and mechanical functions(e.g.,pneumatophores that in

140、flate a sail via synthesis of carbon monoxide)that fuse to appear as a single organism.EI stands poised to leverage these alternative strategies to environmental adaptation to improve maneuverability,agility,and efficiency in achieving tasks.Learning.Many challenges remain when considering how bestt

141、o enable embodied intelligent machines to learn.Traditionally,learning has implied neural plasticity.But,with the construction of robots from increasingly exotic and pliable materials,the bodies of future robots will also likely“learn.”As an example,Hebbian learning has long served as a cornerstone

142、for neural plasticity in AI:synapses strengthen when their pre-and post-synaptic neurons fire together and weaken when they do not.There is a morphological analogue of this process in nature,which is known as Wolffs law:bone strengthens under specific load signatures and weakens under other load con

143、ditions.To date,however,there are few examples of robots built from materials capable of dynamic stiffening and softening.Recently,however,there is an example of a 2D network of motors and flexible beams capable of tuning itself for learning,a true mechanical neural network(Lee et al.,2022).When exa

144、mples like this use materials that can be processed more intricately,the bodies as well as the brains of future robots may be able to learn how best to detect co-occurring features of external challenges(or internal challenges,such as injury),and prepare themselves,morphologically and neurally,to gr

145、apple with those challenges when the re-occur.How EI systems should best transform their bodies in general,and how such change may complement more traditional neural learning,has yet to be determined.Language.Embodied intelligence stands poised to rectify manyof the fundamental problems currently pl

146、aguing non-embodied AI,exemplified by the current state of the art in Large Language Models(LLMs).For example,all LLMs suffer from hallucinations:the generation of non-factual verbal statements.If future embodied intelligences are required to demonstrate,physically,their verbal claims,a self-correct

147、ing mechanism for confabulation becomes possible.Another route to embodied brakes on verbal Figure 4.Examples of adaptation in biology,in comparison to artificial systems.Sources:Fathtabar et al.,2023&dOliveira,202110confabulation could be for embodied intelligences to self-narrate their actions.If

148、a loss function ties those verbal descriptions to those actions,and an LLM is trained on these narrations,there is less likelihood for confabulation as all verbal training data would be factual and physically plausible.However,whether these or other ways for embodiment to constrain LLM pathologies w

149、ill be effective remains an open challenge.Control of the EI systems could initially be more difficult thanthose with a discrete boundary on their bodies and fewer DOFs.Machines with more DOFs than necessary to perform a task are considered kinematically redundant.While offering flexibility,the kine

150、matic redundancy also introduces challenges in choosing the most efficient or smoothest motion.Further,more joints translate to more variables to control.Instructing the robot on how to move each joint in a coordinated way to achieve a specific goal becomes intricate and requires sophisticated progr

151、amming techniques.In EI,however,there is an opportunity to selectively remove DOFs or add more.A Tapestry of ChallengesPerception,Motion,and Adaptation,while we list them as separate pillars,are really an interdependent tapestry that requires concurrent research efforts.One important example of a co

152、mmon thread between them is the challenge of information density.Too little data interpreted from the environment and the system will not have enough information to be considered useful,and too much data and the machine will not be able to interpret the environment to process a state and respond in

153、time to be agile.Perception needs to amplify or filter data,Adaptation needs to understand the resulting information,and Motion needs to adjust for it.Handling the data throughput requires a basis of communication and processing information between the pillars.It is often misconstrued that organisms

154、 are optimal for their environment;however,these ostensibly optimal solutions are usually examples of exaptation.The adaptation of a prior trait for a new function is rarely optimal,but“good enough”for survival.A non-exhaustive set of(good enough)solutions to this problem lies in organismal biology;

155、where animals use at least three approaches:1.Focusing on Relevant Stimuli using(i)Sensory Organs:Each sense organ(eyes,ears,nose)is specialized to detecta specific type of information.This reduces the overall dataintake by focusing on relevant stimuli.For example,an owlshighly sensitive ears allow

156、it to pinpoint prey location inthe dark,filtering out irrelevant visual cues.(ii)BehavioralResponses:Organisms learn to associate specific stimuli withthreats,food,or mates.This approach allows them to focusattention on these important cues and ignore the rest.Forinstance,a bee recognizes the scent

157、of flowers and focuseson following it,filtering out other odors in the environment.2.Responding Based on Needs via(i)Internal State:Anorganisms internal state(hunger,thirst,fear)influences how it interprets sensory information.A hungry animal might prioritizefood-related cues,filtering out others.(i

158、i)Limited Attention:Brains dedicate processing power to the most important tasksat hand.This approach helps filter out less critical informationduring complex situations.For example,a gazelle beingchased by a cheetah will focus on escape routes,filtering outbackground details like potential food sou

159、rces.3.Leveraging Past Experiences by(i)Learning and Memory:Organisms learn from past experiences to identify patternsand predict future events.This allows them to filter outunexpected or irrelevant information.For example,a birdthat has been stung by a brightly colored caterpillar will avoidsimilar

160、ly colored ones in the future.(ii)Innate Responses:Manyorganisms have pre-programmed responses to specific stimuli,filtering out the need to analyze complex information everytime.This phenomenon is often seen in escape reflexes orpredator recognition in young animals.(iii)Remodeling andGrowth:one of

161、 many examples include bones strengthening ofareas where stress is commongetting stronger based on use.By using these strategies,organisms can effectively survive in the real world.They focus on the information crucial for survival and reproduction,filtering out the vast amount of irrelevant data.Fe

162、ature SelectionFeature ExtractionRegularizationChoosing Appropriate AlgorithmsThis involves identifying and discarding irrelevant or redundant features.This can be done through various methods like correlation analysis or using machine learning models for feature importanceThis creates a new set of

163、features,often lower in dimension,that capture the essential information from the original features.Techniques like Principal Component Analysis(PCA)and Autoencoders fall under this category.These techniques penalize models for having too many complex features,encouraging simpler models that are les

164、s prone to overfitting in high dimensions.Some algorithms are more susceptible to the curse than others.For instance,k-Nearest Neighbors(KNN)struggles in high dimensions,while deep learning models can sometimes handle it better.Figure 5.Four stages of feature integration in artificial systems.11 Fur

165、ther ChallengesOur workshop on the Future Directions of Embodied Intelligence identified these key challenges and opportunities related to the synthesis of these pillars.For example,the challenge described in Adaptation describes the problem of controlling high DOFs systems;however,a potential solut

166、ion of growth is described in the Motion pillar.The entangled nature of challenges and solutions is a problem due to the transdisciplinary knowledge requirements.The difficulty in solving this challenge,however,is what motivates the collaboration of research disciplines to solve them.Correspondingly

167、,a huge challenge is to reduce the barriers to knowledge transfer between groups of researchers in EI.Therefore,the workshop participants felt that clear definitions and taxonomy are crucial.Interestingly,LLMs may actually play a crucial role in easing collaboration in this respect.The participants

168、also highlighted the need for metrics to measure progress and standards to ensure consistency.Physically,the discussion on that topic focused on fundamental limits of information rate and energetic limitations of materials.The workshop also identified many successful examples of EI;however,their rel

169、iability issues and the difficulty to manufacture are limiting their utility.The importance of Testing-Evaluation-Validation(TEV)was also made clear throughout discussions.The synthesis of these pillars into a cohesive and global EI program requires transdisciplinary researchers.Materials science,ma

170、nufacturing,computer science,mechanical engineering,and EI design need to be tightly integrated.Researchers need to explore new materials suitable for EI and determine if existing materials can be sufficiently engineered to provide the necessary physical substrate for EI.Further,the workshop emphasi

171、zed fostering a diverse research community and the importance of advancements in energy storage technologies for powering EI systems.12Research OpportunitiesIn the next 20 years,robotics will use EI to better leverage hardware examples from the animal world,interface with the world and within themse

172、lves in increasingly analog fashion,as well as adapt artificial computational approaches to command machines.As new technologies such as Additive Manufacturing(AM),General Pretrained Transformers(GPTs),etc.permit improvements-in and fusion-of Perception,Motion,and Adaptation,the body and brain of au

173、tonomous intelligent machines will become more tightly coupled,blurring the distinction of the two.As EI becomes a ubiquitous and transformative force across various domains,it will reshape daily life,healthcare,manufacturing,and more.OverviewEI improves the compromise between agility(e.g.,accelerat

174、ion and turn radius)and endurance(i.e.,how long it can operate for)in robots.Figure 4 outlined the organismal examples we believe best exemplify the potential of EI.Figure 6,in turn,describes an exciting potential approach to achieving similar capabilities using the coordination of low level,EI enab

175、led,robot modules.Following this high level roadmap,we expect that,in the near term,EI will provide analog sensing,actuation,and computation layers for improved compromises between agility and efficiency on traditional robot bodies with some examples of artificial“organ systems”within these robots.I

176、n the next 10 years,we expect a set of functional modules(akin to organs)to emerge with a basis function that defines their coordination for particular tasks.In the longer term(20 years),we expect that biohybrid modules(including features of muscle,neuron,mycelium,plant cell,etc.)will coordinate int

177、o more complex,synthetic animals,that coordinate to perform more generalized jobs(e.g.,health care,agriculture,disaster relief).Research TrajectoryNear-Term(5 Years)In the immediate future,EI will augment existing robot architectures.These robots,equipped with simple control loops(informed by analog

178、 sensing and processing layers)and commanding actuators(e.g.,continuum,compliant,standard)will be capable of reduced energy expenditure during mobility tasks or more dextrous performance in assembly tasks,for example.These robots may feature,as an example,endoskeletal structures with soft actuators

179、and skins,mediating reconfigurability based on task requirements.The use of compliant manipulators and soft skins will improve their agility and endurance compared to non-EI systems.5-year goals Develop consensus metrics for energy consumption duringstate transitions(e.g.,trotting to cantering),as w

180、ell as agility(e.g.,acceleration and turn radius).Establish foundational control strategies using logical basisfunctions for coordination of tasks.Mid-Term(10 Years)Over the next decade,EI is expected to leverage prior results in analog sense-act-respond functions to produce a set of low-level robot

181、s that demonstrate these principles with specific functions,akin to organs or“polyps”seen in biology(Figure 4).The results may be akin to reconfigurable systems of modules mediated by analog computational layers that can configure for(as an example)external dexterity or(another example)internal oper

182、ational efficiency for existing tasks.Importantly,the development of basis functions for the set of modules will play a critical role in this phase,allowing robots to be dynamically assembled and disassembled in response to environmental or task changes.Near term(5 years)Mid-term(10 years)Long-term(

183、20 years)There is some amount of embodied intelligence layered on traditional physical agents.Functional modules are developed with a set of associated basis functions that describe their potential coordination.Biohybrid modules are able to be configured in order to accomplish generalized job tasks.

184、Figure 6.Visual representations of artificial systems at 5,10,and 20 years years that utilize embodied intelligence.Initially,robots will have layers ofanalog sensing and processing layers that inform simple control loops and a mixture of electric motors and continuum compliant mechanisms.After deca

185、des of research,these layers will become modules that can assemble into more and more complex volumes using algorithms informed by basis functions that coordinate module linkages.1310-year goals Enumeration of agility and endurance requirements forgeneral purpose robotics(these numbers should be arr

186、ivedat beyond just EI community)Define a set of low-level EI modules that address therequirements for agility and endurance Algorithms developed that provide the basis forcoordination between these modules(digital and analogsolutions)Long-Term(20 Years)In the long term,EI researchers will understand

187、 how to best leverage living and synthetic approaches to build low-level EI modules.The basis functions to coordinate the low-level biohybrid robots to autonomously assemble and disassemble themselves into more complex,high-level robots will be known.These high-level robots are more sophisticated,ca

188、pable of performing complex tasks and adapting to changing environments.This synthesis will enable the development of general-purpose robots capable of growth,reconfiguration,and continuous adaptation.Logical basis functions(e.g.,autonomous material computation)(Yamada et al.,2022)will be fully inte

189、grated into the robots architecture,enabling seamless coordination across multiple robots in various environments.In addition to the coordination of low-level robots,we also anticipate autonomous coordination between multiple(and different)high-level robots.20-year goals Develop autonomous material

190、systems(AMS)that allow forindependent sensing and dynamic reconfiguration.Implement neuron-based computing for acceleratedadaptation and coordination of large robot assemblies.Advance multiplexed high-DOF actuator arrays to supportsophisticated motion and structural integrity duringassembly and disa

191、ssembly.Robust approaches to maintaining life in real worldenvironments,as well as mediating their interface withartifices.Communication protocols in addition to radio frequency andvisual spectrum signaling,such as acoustic and chemical.Opportunities to Achieve these GoalsMaterials and Manufacturing

192、Advancements in voxel-based manufacturing,such as Volumetric Additive Manufacturing,(Kelly et al,2019)will enable the creation of multifunctional materials with integrated sensing,actuation,and planning capabilities.Robots will be designed with heterogeneous materials that eliminate the need for dis

193、tinct subsystems,streamlining production and enhancing functionality.Sustainable design principles will lead to robots that grow,reconfigure,and strengthen over time,with minimal environmental impact.These robots will naturally degrade at the end of their life cycle,contributing to a circular econom

194、y.The synthesis of Autonomous Material Systems(AMS)(Howard et al,2019)will fuse sensing-computing-responding to formable elements that allow the construction of EI machinery.AMSs were a prominent discussion in the Motion,Perception,and Adaptation pillarsforming a basis of material science research e

195、ffort where sensing,actuation,and computation become part of a single material element.An early example of an AMS was given as a BelousovZhabotinsky redox reaction inside a thermally swellable gel to maintain a reaction clock speed independent of external temperature conditions(Yamada et al.,2022).S

196、pecific Example of Autonomous MaterialsAutonomous material systems are a class of composites that can independently perform tasks by sensing,processing,and responding to environmental stimuli without external intervention.This capability could accelerate the development of robots with embodied intel

197、ligence,where intelligence is not just a function of computational processing but is distributed throughout the robots body,integrated into its physical structure.Embodied intelligence in robotics refers to the concept that a robots intelligence emerges from the interaction between its body and the

198、environment.Autonomous materials play a critical role in this by enabling the robot to react and adapt at the material level.For example,a robot could be constructed using materials that change shape or stiffness in response to temperature,light,or mechanical stress.These changes could alter the rob

199、ots behavior in real-time,enabling it to navigate complex terrains,avoid obstacles,or even repair itself(an aspect of improved endurance).Robots built using autonomous materials can exhibit a high degree of adaptability and responsiveness to their surroundings.These materials can be designed to poss

200、ess different levels of autonomy,from simple feedforward actions to complex decision-making processes.By embedding intelligence directly into the material(Yamada et al.,2022),robots could operate more efficiently in dynamic environments,reducing the need for centralized control systems.As a guidelin

201、e for EI development using AMS,it is essential to consider the structural complexity and autonomy of the materials used.The framework categorizes materials based on their structural complexity(N)and autonomy(A).For instance,a robot made of N=3 materials(such as composite materials with engineered mi

202、crostructures)and A=3 autonomy(such as smart materials that can sense and actuate)would have a moderate level of embodied intelligence,suitable for tasks like environmental monitoring or exploration in hazardous conditions.By advancing the integration of autonomous materials into robotic systems,it

203、could be possible to build(maybe grow)robots that are more resilient,efficient,and capable of performing tasks in unpredictable and unstructured environments.The evolution of these systems could lead to robots that are not only more independent but also more harmonious in their interaction with the

204、world,much like biological organisms.14 The development of AMS for embodied intelligence in robotics is still in its infancy.However,as our understanding of these materials deepens,and as manufacturing techniques improve,we can expect to see a new generation of robots that are smarter,more adaptable

205、,and capable of undertaking tasks that were previously unimaginable.The intersection of materials science,robotics,and artificial intelligence will be the driving force behind this innovation,leading to a future where robots with embodied intelligence become a vital part of our technological landsca

206、pe.Adaptation and ComputationFuture EI systems will feature hyper-embedded computation that integrates seamlessly with physical substrates,enabling more efficient and adaptive behaviors.Moving beyond digital,a return to analog and biotic computation will enhance response speeds and/or reduce power c

207、onsumption significantly.Robots will develop unified perception and motion capabilities,allowing them to adapt actively and autonomously to diverse environments.Decentralized adaptation mechanisms could aid in cohesive sensing and actuation,making robots more responsive and versatile.As GPTs become

208、more sophisticated and allow for better Large Language Models(LLMs),as well as other interactions modes(perhaps a Large Touch Model or Large Smell Model),the communication between Perception,Adaptation,and Motion domains will become more efficient and capable.Independent robot modules can take advan

209、tage of EI layers representing GPTs physically,could operate as swarms,high level assemblies,and swarms of high-level assemblies their fusion guided by a set of basis functions defined as these modules converge on a set of low-level functions(akin to organ systems).Specific Example of Basis Function

210、s for the Coordination of Organ ModulesEmbodied Intelligence(EI)presents a promising avenue for advancing the coordination of low-level robot modules with high-level robotic interfaces in complex environments.Before proposing a basis function that could enhance such coordination,its essential to fir

211、st consider the inherent complexities involved in the process of robotic self-assembly.One of the primary challenges is ensuring module compatibility.Each module must have interfaces that are not only compatible for attachment but also capable of facilitating the seamless transfer of power and data.

212、Furthermore,communicationbetween modules is critical.Each module needs a reliable method to communicate its identity,current status,and desired configuration with other modules to enable coordinated assembly.Another critical consideration is the availability of an energy source.Without a steady powe

213、r supply,modules cannotactivate or move,rendering the assembly process impossible.Additionally,environmental factors such as temperature,gravity,and other external conditions can significantly influence the assembly process,adding another layer of complexity.Given these challenges,a potential soluti

214、on lies in developing a multi-dimensional potential field as a basis function.This basis function would integrate several key components:Geometric Compatibility:This component would accountfor the shape and size of each module,as well as theconfiguration of their attachment points,ensuring thatmodul

215、es can physically connect with one another.Functional Compatibility:This would represent thecapabilities of each module,such as sensing,actuation,or computation,allowing for the creation of functionallycomplementary assemblies.Communication Protocol:This component would definethe method and format o

216、f data exchange between modules,ensuring that they can effectively communicate andcoordinate their actions.Energy State:This would monitor the energy level of eachmodule,ensuring that modules with adequate energy areprioritized in the assembly process.Environmental Factors:This component would inclu

217、deparameters for external conditions like temperatureand gravity,allowing the system to adapt to varyingenvironments.The operation of the potential field would hinge on several interaction principles.Modules with compatible geometric and functional interfaces would experience an attractive force,dra

218、wing them together.Conversely,modules with incompatible interfaces or overlapping volumes would be subject to a repulsive force,preventing erroneous connections.Communication between modules would allow them toexchange vital information regarding their status and desired configuration,directly influ

219、encing the dynamics of the potential field.Moreover,modules would seek to optimize their positions in the assembly to maximize energy efficiency,potentiallyextending the operational lifespan of the robotic system.In response to changing environmental conditions,the modules would exhibit environmenta

220、l adaptation,adjustingtheir behavior to ensure successful assembly despite external challenges.Additional considerations include the necessity for a dynamic potential field.The field must be capable of evolving in real-time to accommodate changes in the environment or the configuration of the module

221、s.Furthermore,an error correction mechanism would be essential to handle issues such asmisaligned modules,ensuring that the assembly process can recover from mistakes.For more complex robotic systems,hierarchical assemblymight be required.In such cases,modules would first form subassemblies before b

222、eing integrated into the final system.A higher-level control system could oversee this process,providingoverall guidance and coordination to ensure the successful completion of the assembly.15However,several challenges remain.Computational complexitycould become a significant issue,as calculating an

223、d updating the potential field for a large number of modules might be resource intensive.There is also the risk of the assembly process becoming trapped in local minima,resulting in suboptimalconfigurations.Additionally,physical constraints such as frictionand elasticity could interfere with the int

224、ended assembly process.Inspiration for overcoming these challenges could be drawn from natural systems of self-assembly,such as the formation of crystals or the behavior of biological cells.By studying these systems,valuable insights might be gained into the development of robust basis functions and

225、 optimization strategies.In summary,by carefully addressing these considerations,it is possible to develop a sophisticated basis function that enables the self-assembly of robots with increasing complexity.Leveraging Embodied Intelligence(EI)in this context could allow robots to achieve greater auto

226、nomy and adaptability,significantly enhancing their ability to interact with and respond to dynamic environments.This focused research trajectory could lead to groundbreaking advancements in the field of robotics,particularly in the assembly and disassembly of modular robotic systems.By focusing on

227、the use of Embodied Intelligence to coordinate the assembly and disassembly of low-level robots through logical basis functions,this research trajectory aims to advance the field of robotics significantly.Achieving these goals will require continued innovation in materials science,AI,and manufacturi

228、ng techniques,as well as a strong commitment to interdisciplinary collaboration and policy development.The successful integration of EI into robotic systems will pave the way for more agile,adaptable,and autonomous machines capable of meeting the demands of a rapidly changing world.Application Focus

229、 AreasDaily Life and Labor ReplacementEI will be deeply embedded in everyday appliances,dramatically reducing the time and effort required for routine tasks.For instance,robotic cleaners will save hours each week,while wearable exosuits will enhance human physical capabilities.Biohybrid robots,mimic

230、king the responsiveness and efficiency of animals,will assist with various tasks.These advancements will address labor shortages in developed regions by automating tasks currently performed by humans due to cost advantages or need for safe decision-making and motion adaptability,especially in danger

231、ous conditions.Healthcare and RoboticsThe healthcare sector will see a significant influx of soft robots designed for patient transfer and rehabilitation.These robots will be more affordable and accessible,driven by advances in soft robotics and biohybrid designs.This will alleviate the strain on he

232、althcare systems and improve patient care.Soft,adaptable robots will offer gentle and precise assistance,enhancing recovery and comfort for patients,while EI-enabled robotic prosthetics or exo-skeletons will greatly augment the quality of life.They can also allow or facilitate emergency responses in

233、 hazardous situations.Advanced Task-Specific RobotsInnovative robots like personal assistants capable of retrieving objects and self-cleaning will become commonplace.These ultra-low-cost robots will handle specific tasks efficiently and adapt to disturbances,offering practical solutions for everyday

234、 problems.This new generation of robots will be simpler yet more effective,reflecting a shift towards task-specific designs that prioritize functionality and cost-efficiency.How to Accelerate the FieldOvercoming time-scale challenges in additive manufacturing and leveraging new 3D printing technolog

235、ies for example,will be vital.Enhanced design tools will allow early specification of goals and intents,streamlining the manufacturing process.Competitions akin to DARPA challenges will foster innovation and interdisciplinary cooperation,while funding for basic research and collaborative projects wi

236、ll drive continuous progress.Industry partnerships will provide access to cutting-edge technologies,and incentives will encourage the establishment of robotics departments and tenure opportunities for young researchers.Collaboration and CommunityWithin two decades,EI will transform from a niche rese

237、arch area to a foundational technology embedded in all aspects of life.The integration of advanced computation,adaptive materials,and interdisciplinary collaboration will lead to smarter,more efficient,and more versatile robots.These advancements will not only improve daily life and healthcare but a

238、lso drive new industries and economic growth,marking a new era of human-robot interaction and collaboration and laying the foundation for an ubiquitous robotic society.Interdisciplinary collaboration will be crucial,with AI/ML researchers working alongside soft robotics designers,engineers and mater

239、ial scientists to achieve specific goals,such as building smart soft-legged robots.Training programs will evolve to prepare the next generation of EI researchers,emphasizing the integration of robots into larger systems.Common platforms for materials synthesis and machine learning will facilitate ba

240、rrier-free collaboration,driving innovation and efficiency.Metrics and Evaluation A significant challenge in advancing EI is the lack of standardized metrics for evaluating progress in assembly and disassembly tasks.To address this,the establishment of an EI leaderboard(perhaps an online tool with o

241、versight that tracks metrics in important categories,such as agility,endurance,and other critical attributes.Creating compelling challenges,similar to the DARPA Driverless Car challenge,will focus research efforts and inspire innovation in the field.16 The DARPA Robotics Challenge for Embodied Intel

242、ligence would push the frontiers of robotics by focusing on the deep integration of Embodied Intelligence with existing robots(e.g.,deformable,sensing skins overlaid onto a quadruped like Spot).This competition would require participants to design robots that can seamlessly interact with dynamic,uns

243、tructured environments,emphasizing the synergy between a robots body and its decision-making processes.In this challenge,robots would be tasked with completing complex,real-world scenarios such as disaster response or search and rescue missions.These tasks would demand not only physical robustness b

244、ut also the ability to process sensory information in real-time,make autonomous decisions,and adapt to unforeseen circumstances.The essence of embodied intelligence lies in how the robots physical form and sensory inputs are tightly interwoven with its cognitive functions,enabling more natural,fluid

245、 interactions with the environment.For example,the competition would stress autonomy,with robots expected to navigate difficult terrains,manipulate objects,and interact with humans in cooperative tasks.Success would depend on the robots ability to integrate learning and adaptability,improving perfor

246、mance as it encounters new challenges.Resilience would also be key,as robots must demonstrate the ability to operate under adverse conditions,recover from disruptions,and self-diagnose issues.Ultimately,the most successful robots would perform agile tasks at the lowest energy expenditures.Through th

247、is challenge,DARPA would drive advancements in embodied intelligence,encouraging the development of robots that are not only mechanically capable but also exhibit the adaptive,responsive behaviors necessary for real-world applications.17ConclusionThe workshop identified several key challenges and ga

248、ps that need to be addressed to make EI a critical component in the development of intelligent autonomous machines,including performance metrics and benchmarks.Some of these benchmarks include data rates for useful information processing,power,efficiency,and controllable degrees of freedom(DOFs).Eve

249、n in these simple benchmarks,there are complex tradeoffs.For example,software processing benefits from reduced DOFs while additional DOFs from more complex hardware provides more maneuverability.Additionally,standards for robot components and their interconnectivity need to be developed alongside lo

250、ng-term reliability testing protocols(i.e.T-E-V)to ensure these systems can perform consistently over extended periods.Technological and material advances play a critical role in the development of EI.Identifying and developing new materials suitable for EI,understanding their fundamental chemistry

251、and physics,and leveraging advanced manufacturing techniques like volumetric printing,biohybrids,and AMS.Additionally,ensuring the reliability of EI systems through prolonged testing and real-world deployments is crucial.Developing systems that can continuously learn and adapt through interaction wi

252、th their environment is equally important;transferring learning from one agent to the next is also a critical component to these systems.Energy challenges,particularly regarding capacity and efficiency,must be addressed.Treating EI as a dynamic,adaptive process rather than a static state and designi

253、ng systems that can adapt over time are key considerations.Integrating materials development with the design and manufacturing of EI systems is essential,requiring innovative paradigms to support dynamic and adaptable embodiments.Balancing algorithmic and hardware integration is another significant

254、challenge.It is important to optimize the roles of physical and digital algorithms and hardware to achieve the best EI results.Exploring neuromorphic computing and low-energy embedded intelligence can offer new solutions.Using AI to enhance the design and functionality of EI systems will also be ben

255、eficial.Community cohesion and interdisciplinary collaboration are other significant areas requiring attention.There is a need to enhance cooperation across AI,materials science,manufacturing,and robotics communities,addressing divides between different research sectors and encouraging interdiscipli

256、nary discussions at conferences and other forums.Improving communication and idea-sharing within the community is vital for progress.In summary,building a cohesive community,innovating in material science and energy solutions,implementing long-term and adaptive testing,and maintaining a strong inter

257、disciplinary approach are all necessary steps to address the current challenges and advance the field of Embodied Intelligence.This workshop was a first step in building that community with researchers from US and South Korea mapping the future.18 BibliographyPfeifer,R.,&Bongard,J.(2006).How the Bod

258、y Shapes the Way We Think:A New View of Intelligence.MIT Press.Wikipedia contributors.(2024,June 17).Lesser blind mole-rat.Wikipedia.https:/en.wikipedia.org/wiki/Lesser_blind_mole-ratWikipedia contributors.(2024a,May 30).Mexican tetra.Wikipedia.https:/en.wikipedia.org/wiki/Mexican_tetra Wikipedia co

259、ntributors.(2024c,July 11).Kauai cave wolf spider.Wikipedia.https:/en.wikipedia.org/wiki/Kaua%CA%BBi_cave_wolf_spider Laurent,G.,Lacoste,D.,&Gaspard,P.(2021).Emergence of homochirality in large molecular systems.Proceedings of the National Academy of Sciences,118(3),e2012741118.Kirstines.Dk.(2016,Ju

260、ne 14).South America Animals.Pinterest.https:/ enduring autonomous robots via embodied energy.Nature,602(7897),393402.https:/doi.org/10.1038/s41586-021-04138-2Fathtabar,A.,Ebrahimzadeh,A.,&Kazemitabar,J.(2023).Ant path integration:a novel optimization algorithm inspired by the path integration of de

261、sert ants.Neural Computing&Applications,35(23),1729317318.https:/doi.org/10.1007/s00521-023-08611-zdOliveira,R.(2021,February 6).Portuguese man of war.Its a colony of hundreds of polyps of different types-Awesome.Pinterest.https:/ neural networks:Architected materials that learn behaviors.Science Ro

262、botics,7(71).https:/doi.org/10.1126/scirobotics.abq7278Howard,D.,Eiben,A.E.,Kennedy,D.F.et al.Evolving embodied intelligence from materials to machines.Nat Mach Intell 1,1219(2019).https:/doi.org/10.1038/s42256-018-0009-9 Yamada,Y.,Ito,H.,&Maeda,S.(2022).Artificial temperature-compensated biological

263、 clock using temperature-sensitive BelousovZhabotinsky gels.Scientific Reports,12(1).https:/doi.org/10.1038/s41598-022-27014-zKelly,B.E.,Bhattacharya,I.,Heidari,H.,Shusteff,M.,Spadaccini,C.M.,&Taylor,H.K.(2019).Volumetric additive manufacturing via tomographic reconstruction.Science,363(6431),107510

264、79.https:/doi.org/10.1126/science.aau711419Appendix I Workshop Attendees Workshop Co-chairsJoshua Bongard,University of VermontKyujin Cho,Seoul National UniversityYong-Lae Park,Seoul National UniversityRobert Shepherd,Cornell UniversityWorkshop Participants Cameron Aubin,University of MichiganJoonbu

265、m Bae,Ulsan National Institute of Science and TechnologyJosh Bongard,University of VermontMark Campbell,Cornell UniversityKyujin Cho,Seoul National UniversityDavid Hu,Georgia Tech UniversityAyoung Kim,Seoul National UniversityDaekyum Kim,Korea UniversityHyoun Jin Kim,Seoul National UniversityDaniel

266、E.Koditschek,University of PennsylvaniaJesung Koh,Ajou UniversitySeung Hwan Koh,Seoul National UniversityKi-Uk Kyung,The Korea Advanced Institute of Science and TechnologyJeffrey Lipton,Northeastern UniversityRob MacCurdy,Colorado UniversityLakshminarayanan Mahadevan,Harvard UniversityFrank Park,Seo

267、ul National UniversityYong-Lae Park,Seoul National UniversityDaniela Rus,Massachusetts Institute of TechnologyJee-Hwan Ryu,The Korea Advanced Institute of Science and TechnologyRobert Shepherd,Cornell UniversityDongjun Shin,Yonsei UniversityJeong-Yun Sun,Seoul National UniversityMichael Tolley,Unive

268、rsity of California,San DiegoRyan Truby,Northwestern UniversityTJ Wallin,Massachusetts Institute of TechnologyVictoria Webster-Wood,Carnegie Melon UniversityJinkyu Yang,Seoul National UniversityGovernment Observers Jean-Luc Cambier,OUSD(R&E)Basic Research OfficeDokeun Kang,Agency for Defense Develop

269、mentDai Hyun Kim,OUSD(R&E)/OASD(CT)/TAIAKiho Kwak,Agency for Defense DevelopmentWook-hyang Kwon,Ministry of Trade,Industry and EnergyJihong Min,Agency for Defense DevelopmentHyeong-tae Park,Ministry of Trade,Industry and EnergyHyunsoo Woo,Ministry of Trade,Industry and EnergyVT-ARC TeamKate Klemic,V

270、irginia Tech Applied Research CorporationJae Kwak,Virginia Tech Applied Research CorporationSean Lemkey,Virginia Tech Applied Research CorporationSithira Ratnayaka,Virginia Tech Applied Research Corporation20 Participant Short BiographyCameron AubinAssistant Professor,University of Michigancaubinumi

271、ch.edu|https:/ Aubin is an assistant professor in the Robotics Department at the University of Michigan.He previously conducted his graduated work at Cornell University,where he received his Ph.D.in Mechanical Engineering in 2023.Camerons work centers on improving the endurance,adaptability,and auto

272、nomy of robots through the integration of multifunctional,biologically-inspired energy systems.His interests also include soft robotics,microrobotics,and advanced materials and manufacturing.He has published in a number of reputable journals,including Nature and Science,and his research has been fea

273、tured by several popular media outlets,including CNN,PBS,BBC,Wired,New Scientist,ARS Technica,and more.Joonbum BaeProfessor,Ulsan National Institute of Science and Technologyjbbaeunist.ac.kr|http:/birc.unist.ac.krJoonbum Bae is Professor of Department of Mechanical Engineering and the Director of Bi

274、o-Robotics and Control(BiRC)Lab of Ulsan National Institute of Science and Technology(UNIST).He earned his B.S.degree in mechanical and aerospace engineering from Seoul National University,followed by M.S.and Ph.D.degrees in mechanical engineering,along with an M.A.in statistics,from the University

275、of California,Berkeley.His current research interests include modeling,design,and control of human-robot interaction systems,soft robotics,and biologically inspired robot systems.Additionally,he is a CEO and founder of a startup Feel the Same,Inc.,which develops wearable soft sensor systems.Recogniz

276、ed for his academic achievements,he was appointed as a Rising-Star Distinguished Professor of UNIST.He has received prestigious awards including the Samsung Scholarship for his Ph.D.studies,the Young Researcher Award from the Korea Robotics Society,the Korean Government Minister Awards from the Mini

277、stry of Public Safety and Security and the Ministry of Science,ICT and Future Planning,Best Teaching Award from UNIST,and the Grand Prize from the UNIST Outstanding Faculty Awards.He led the Team UNIST at$10M ANA Avatar XPRIZE,which is a global avatar robot competition,achieving a sixth place in the

278、 finals.Josh BongardProfessor,University of Vermontjbongarduvm.edu|https:/www.meclab.org/Josh Bongard is the Veinott Professor of Computer Science at the University of Vermont and director of the Morphology,Evolution&Cognition Laboratory.His work involves automated design and manufacture of soft-,ev

279、olved-,and crowdsourced robots,as well as AI-designed organisms.A PECASE,TR35,and Cozzarelli Prize recipient,he has received funding from NSF,NASA,DARPA,ARO and the Sloan Foundation.He is the co-author of the book How The Body Shapes the Way We Think,the instructor of a reddit-based evolutionary rob

280、otics MOOC,and director of the robotics outreach program Twitch Plays Robotics.Mark CampbellProfessor,Cornell Universitymc288cornell.edu|http:/campbell.mae.cornell.edu/Mark Campbell is the John A.Mellowes Professor of Mechanical&Aerospace Engineering at Cornell University.He received his B.S.in Mech

281、anical Engineering from Carnegie Mellon,and his M.S.and Ph.D.in Aeronautics and Astronautics from MIT.His research interests are in the areas of autonomous systems including robots,self-driving cars,UAVs and spacecraft,with a focus on algorithms and hardware verification including estimation,machine

282、 learning,perception,sensor fusion,planning under uncertainty,multi-agent systems,and human-robotic teaming and decision making.Professor Campbell has led multiple collaborative research grants with DARPA,AFOSR,ARO,ONR and NSF,including leading Cornells DARPA Urban Challenge self-driving car team,on

283、e of six finishers of the race.He also served as a member of the U.S.Air Force Science Advisory Board,21advising leadership on science,technology and investments,reviewing research labs,and leading a study on Unintended Behaviors of Autonomy.For his work with the board,he received the U.S.Air Force

284、Chief of Staff Award for Exceptional Public Service,the highest-level award granted by the U.S.Air Force to non-employee civilians.Prof.Campbell is a Fellow of the IEEE,AIAA and ASME.Kyujin ChoProfessor,Seoul National Universitykjchosnu.ac.kr|https:/www.biorobotics.snu.ac.kr/lab-membersKyu Jin Cho i

285、s a Professor of Mechanical Engineering and the Director of Soft Robotics Research Center and Biorobotics Lab at Seoul National University.He received his Ph.D.in mechanical engineering from MIT and his B.S and M.S.from Seoul National University.He was a post-doctoral fellow at Harvard Microrobotics

286、 Laboratory before joining SNU in 2008.He has been exploring novel soft bio-inspired robot designs,including a water jumping robot,various shape changing robots and soft wearable robots for the disabled.He has received the 2014 IEEE RAS Early Academic Career Award for his fundamental contributions t

287、o soft robotics and biologically inspired robot design.He has published a Science paper on water jumping robot and several papers in Science Robotics with novel robot designs.He served RAS as associate VP of Publication Activities Board for six years,and is currently serving RAS as Vice President of

288、 the Technical Activities Board.David HuProfessor,Georgia Tech Universityhume.gatech.edu|https:/www.me.gatech.edu/faculty/huDr.David Hu is Professor of Mechanical Engineering and Biology and Adjunct Professor of Physics at Georgia Institute of Technology.He earned degrees in mathematics and mechanic

289、al engineering from M.I.T.and was a National Science Foundation(NSF)Postdoctoral Fellow at New York University.He isa recipient of the APS Fellowship,the Ig Nobel Prize in Physics(twice),the NSF CAREER award,andthe American Institute of Physics Science Communication Award.He sits on the editorial bo

290、ards ofProceedings of the Royal Society B and Journal of Experimental Biology.He is the author of two books“How to walk on water and climb up walls,”(Princeton University Press)and“The P Word”(Science,Naturally).He lives with his wife and two children in Atlanta,Georgia.Daekyum KimProfessor,Korea Un

291、iversitydaekyumkorea.ac.kr|https:/ Kim received his B.S.degree in Mechanical Engineering from the University of California,Los Angeles,(Los Angeles,CA,USA),in 2015.He earned his Ph.D.degree in Computer Science at KAIST(Daejeon,Republic of Korea),in 2021.He was a Postdoctoral Research Fellow at the J

292、ohn A.Paulson School of Engineering and Applied Sciences,Harvard University(Cambridge,MA,USA),co-affiliated with Wyss Institute.Since September 2023,he has been an Assistant Professor with the School of Smart mobility and the School of Mechanical Engineering,Korea University(Seoul,Republic of Korea)

293、.His research interests are in the areas of machine learning,computer vision,robotics,and digital healthcare.Hyoun Jin KimProfessor,Seoul National Universityhjinkimsnu.ac.kr|https:/aerospace.snu.ac.kr/en/about/faculty?mode=view&profidx=9H.Jin Kim is Professor/Chair in Aerospace Engineering at Seoul

294、National University.She received MSc andPhD degrees from the University of California,Berkeley and BS from Korea Advanced Institute of Scienceand Technology(KAIST),Korea,all in Mechanical Engineering.Her research is on navigation,control andplanning of autonomous robotic systems ranging from ground

295、to flying robots.She has served on theeditorial board of several journals and conferences including IEEE Transactions on Robotics,Mechatronics,an International Journal of IFAC,International Journal of Robotics Research,and IEEE Conference on Robotics and Automation,IEEE/RSJ International Conference

296、on Intelligent Robots and Systems.She is a member of National Academy of Engineering of Korea.22 Ayoung KimAssociate Professor,Seoul National Universityayoungksnu.ac.kr|https:/me.snu.ac.kr/en/snu_professor/kim-ayoung/Ayound Kim is currently working as an associate professor in the department mechani

297、cal engineering at SNU since 2021 Sep.Before joining SNU,she was at the civil and environmental engineering,Korea Advanced Institute of Science and Technology(KAIST)from 2014 to 2021.Dr.Kim has earned both a B.S.and M.S.degree in mechanical engineering from SNU in 2005 and 2007,and a M.S.degree in e

298、lectrical engineering and a Ph.D.degree in mechanical engineering from the University of Michigan(UM),Ann Arbor,in 2011 and 2012.Daniel E.KoditschekProfessor,University of Pennsylvaniakodseas.upenn.edu|https:/kodlab.seas.upenn.edu/Research in Daniels group is focused on the application of dynamical

299、systems theory to the design,construction and empirical testing of machines that juggle,run,climb,and in general,interact physically with their environment to perform useful work.Dan and his group seek to probe the foundations of autonomous robotics by reasoning formally about mathematical models th

300、at represent the successes and limitations of their physical platforms.They maintain close collaborations with biologists,whose insights about animal mobility and dexterity inspire their thinking and designs.Seung Hwan KohProfessor,Seoul National Universitymaxkosnu.ac.kr|https:/me.snu.ac.kr/en/snu_p

301、rofessor/ko-seung-hwan/Seung Hwan Koh is a Professor at Seoul National University,working in the Applied Nano and Thermal Science(ANTS)Lab.Dr.Koh received a PhD in Mechanical Engineering from the University of California,Berkeley in 2006Jesung KohAssociate Professor,Ajou Universityjskohajou.ac.kr|ht

302、tps:/ Koh is an Assistant Professor at Ajou University in the Department of Mechanical Engineering.Dr.Koh received a PhD from Seoul National University in 2014.Dr.Kohs research interests include:Printable&Origami Inspired Robotics,Biologically Inspired Designs&Mechanisms for Robotic Applications,Mic

303、ro Robots Based on the Fiber Reinforced Composite,Microfabrication&Assembly,and Artificial Muscle Actuators(e.g.Shape Memory Alloy actuators).Ki-Uk KyungAssociate Professor,The Korea Advanced Institute of Science and Technologykyungkukaist.ac.kr|https:/irobot.kaist.ac.kr/bbs/content.php?co_id=profes

304、sorKi-Uk Kyung received BS,MS,and Ph.D.degrees in mechanical engineering from the Korea Advanced Institute of Science and Technology(KAIST)in 1999,2001,and 2006,respectively.In 2006,he joined the Electronics and Telecommunications Research Institute and had been the Director of the Smart UI/UX Devic

305、e Research Section.He had been a co-chair of the IEEE Technical Committee on Haptics(TCH)from 2018 to 2021.He received the IEEE TCH Early Career Award in 2015 and the Academic Career Award at Active Materials and Soft Mechatronics 2019.He is currently an associate professor of Mechanical Engineering

306、 and the director of the Human-Robot Interaction Research Center at KAIST,and adjunct professor of Tandon School of Engineering at New York University.His research interests are soft sensors and actuators,haptics,soft robots,and human-robot interaction.23Jeffrey LiptonAssistant Professor,Northeaster

307、n Universityj.liptonnortheastern.edu|https:/Jeff Liptons current work is currently focused on 3D printing and robotics.He focuses on how we can make torque responsive metamaterials and how we can leverage them to make systems with mechanical intelligence.His past work on 3D printed foods and 3D prin

308、ting for the hospitality industry has influenced two of the largest 3D printing companies in America and garnered media attention from the New York Times,BBC,and others.He was the lead developer for the FabHome project which supported life science and food science researchers 3D printing needs on al

309、l six habitable continents.Robert MacCurdyAssistant Professor,Colorado Universitymaccurdycolorado.edu|https:/www.matterassembly.org/Dr.Robert MacCurdy is an assistant professor in Mechanical Engineering(also by courtesy in CS and ECEE)at the University of Colorado Boulder where he leads the Matter A

310、ssembly Computation Lab(MACLab).He is developing new algorithms,materials,and fabrication tools to automatically design and manufacture electromechanical systems,with a focus on robotics.Rob did his PhD work with Hod Lipson at Cornell University and his postdoctoral work at MIT with Daniela Rus.He h

311、olds a B.A.in Physics from Ithaca College,a B.S.in Electrical Engineering from Cornell University,and an M.S.and PhD in Mechanical Engineering from Cornell University.Lakshminarayanan MahadevanProfessor,Harvard Universitylmahadevg.harvard.edu|https:/softmath.seas.harvard.edu/Lakshminarayanan Mahadev

312、an FRS is an Indian-American scientist.He is currently the Lola England de Valpine Professor of Applied Mathematics,Organismic and Evolutionary Biology and Physics at Harvard University.His work centers around understanding the organization of matter in space and time(that is,how it is shaped and ho

313、w it flows,particularly at the scale observable by the unaided senses,in both physical and biological systems).Mahadevan is a 2009 MacArthur Fellow.Frank ParkProfessor,Seoul National Universityfcpsnu.ac.kr|http:/robotics.snu.ac.krFrank C.Park is Professor of Mechanical Engineering at Seoul National

314、University.He received the B.S.in EECS from MIT in 1985,the Ph.D.in applied mathematics from Harvard in 1991,and was on the faculty of the University of California,Irvine from 1991 to 1994.He is a fellow of the IEEE,and has held adjunct faculty positions with the HKUST Robotics Institute in Hong Kon

315、g,the Interactive Computing Department at Georgia Tech,and the NYU Courant Institute.His research interests include robotics,computer vision,mathematical data science,and related areas of applied mathematics.He is a former Editor-in-Chief for the IEEE Transactions on Robotics,developer of the EDX co

316、urse Robot Mechanics and Control I-II,and author(with Kevin Lynch)of the textbook Modern Robotics:Mechanics,Planning,and Control(Cambridge University Press,2017).He served as president of the IEEE Robotics and Automation Society(2022-2023),and is a founder and CEO of the industrial AI startup Saige(

317、http:/saige.ai).Research Interests:Robot mechanics,planning and control;mathematical systems theory;machine learning and mathematical data science;computer vision;related areas of applied mathematics.24 Yong-Lae ParkProfessor,Seoul National Universityylparksnu.ac.kr|https:/me.snu.ac.kr/en/snu_profes

318、sor/park-yong-lae/Yong-Lae Park is Professor in the Department of Mechanical Engineering at Seoul National University(SNU)(2016present).Prof.Park completed his Ph.D.degree in Mechanical Engineering at Stanford University(2010).Prior to joining SNU,he was Assistant Professor in the Robotics Institute

319、 at Carnegie Mellon University(20132017)and Technology Development Fellow in the Wyss Institute for Biologically Inspired Engineering at Harvard University(20102013).His current research interests include artificial skins and muscles,soft robots,wearable robots,medical robots,and inflatable robots.H

320、e received the Best Application Paper Award from the IEEE Transactions on Haptics(2020),the Best Conference Paper Award in the IEEE International Conference on Soft Robotics(2019),Okawa Foundation Research Grant Award(2014),the Best Paper Award from the IEEE Sensors Journal(2013),the NASA Tech Brief

321、 Award(2012).His papers on soft artificial muscles and skin were selected as cover articles in various journals,including Soft Robotics,Advanced Intelligent Systems and the IEEE Sensors Journal,and his work on soft robots were featured in media,including Nature,Discovery News,New Scientist,PBS NOVA,

322、and Reuters.Daniela RusProfessor,Massachusetts Institute of Technologyruscsail.mit.edu|https:/www.csail.mit.edu/person/daniela-rusDaniela Rus is the Andrew(1956)and Erna Viterbi Professor of Electrical Engineering and Computer Science,Director of the Computer Science and Artificial Intelligence Labo

323、ratory(CSAIL)at MIT,and Deputy Dean of Research in the Schwarzman College of Computing at MIT.Prof.Russ research interests are in robotics and artificial intelligence.The key focus of her research is to develop the science and engineering of autonomy.Prof.Rus served as a member of the Presidents Cou

324、ncil of Advisors on Science and Technology(PCAST)and on the Defense Innovation Board.She is a senior visiting fellow at MITRE Corporation.Prof.Rus is a MacArthur Fellow,a fellow of ACM,IEEE,AAAI and AAAS,a member of the National Academy of Engineering,and of the American Academy of Arts and Sciences

325、.She is the recipient of the Engelberger Award for robotics,the IEEE RAS Pioneer award,Mass TLC Innovation Catalyst Award,and the IJCAI John McCarthy Award.She earned her PhD in Computer Science from Cornell University.Robert ShepherdProfessor,Cornell Universityrfs247cornell.edu|http:/orl.mae.cornel

326、l.eduRobert Shepherd is an associate professor at Cornell University in the Sibley School of Mechanical&Aerospace Engineering.He received his B.S.(Material Science&Engineering),Ph.D.(Material Science&Engineering),and M.B.A.from the University of Illinois.At Cornell,he runs the Organic Robotics Lab(O

327、RL:http:/orl.mae.cornell.edu),which focuses on using methods of invention,including bioinspired design approaches,in combination with material science and mechanical design to improve machine function and autonomy.We rely on new and established synthetic approaches for soft material composites that

328、create new design opportunities in the field of robotics.He is the recipient of an Air Force Office of Scientific Research Young Investigator Award,an Office of Naval Research Young Investigator Award,is a Senior Member of the National Academy of Inventors,and his labs work has been featured in popu

329、lar media outlets such as the BBC,Discovery Channel,and PBSs NOVA documentary series.He is an advisor to the American Bionics Project(americanbionics.org)which aims to make wheelchairs obsolete.He is also the co-founder of the Organic Robotics Corporation,which aims to digitally record the tactile i

330、nteractions of humans and machines with their environment.Dongjun ShinProfessor,Yonsei Universitydj.shinyonsei.ac.kr|https:/hcr.yonsei.ac.kr/people.htmlDongjun Shin is a member of the Department of Mechanical Engineering at Yonsei University,working on soft wearables and actuators.Dr.Shin received a

331、 PhD in Mechnical Engineering from Stanford University.25Jeong-Yun SunProfessor,Seoul National Universityjysunsnu.ac.kr|https:/mfsm.snu.ac.kr/index.htmJeong-Yun Sun is currently a professor in the Department of Materials Science and Engineering at Seoul National University(SNU),Republic of Korea.He

332、got his B.S.(2005),M.S.(2007)and Ph.D.(2012)in Materials Science and Engineering at Seoul National University.During his Ph.D.,he had stayed at Harvard University for 4 years as a visiting student.After getting Ph.D.(2012),he started to work as a postdoctoral fellow in School of Engineering and Appl

333、ied Sciences at Harvard University.After his Post-Doc.,he came back to SNU and worked as an assistant professor and an associate professor.His research was focused on developing soft and ionic materials.Based on the materials,he is developing many ionic devices such as sensors,actuators,energy harvesters etc.Dr.Sun has published many high impact peer-reviewed journal papers including Nature,Scienc