SCIENCE ROBOTICS | PERSPECTIVE

ROBOTICS Copyright © 2018 The Authors, some rights reserved; The grand challenges of Science Robotics exclusive licensee Guang-Zhong Yang,1* Jim Bellingham,2 Pierre E. Dupont,3 Peer Fischer,4,5 Luciano Floridi,6,7,8,9,10 American Association 11 12,13 14 15 1 for the Advancement Robert Full, Neil Jacobstein, Vijay Kumar, Marcia McNutt, Robert Merrifield, of Science. No claim 16 17,18 7,8,9 19 Bradley J. Nelson, Brian Scassellati, Mariarosaria Taddeo, Russell Taylor, to original U.S. 20 21 22,23 Manuela Veloso, Zhong Lin Wang, Robert Wood Government Works

One of the ambitions of Science Robotics is to deeply root robotics research in science while developing novel robotic platforms that will enable new scientific discoveries. Of our 10 grand challenges, the first 7 represent underpin- ning technologies that have a wider impact on all application areas of robotics. For the next two challenges, we have included social robotics and medical robotics as application-specific areas of development to highlight the substantial societal and health impacts that they will bring. Finally, the last challenge is related to responsible in- novation and how ethics and security should be carefully considered as we develop the technology further. Downloaded from

INTRODUCTION great strides in realizing its mission of power­ ing components into synthetic structures to Just over a year ago, we published the first issue ing the world’s robots, from space robot chal­ create robots that perform like natural systems. of Science Robotics. Even within this relatively lenges to autonomous driving, industrial (iii) New power sources, battery technol- short period of time, remarkable progress has assembly, and surgery. ogies, and energy-harvesting schemes for long-​ been made in many aspects of robotics—from Given all these advances, what does the ­lasting operation of mobile robots. http://robotics.sciencemag.org/ micromachines for biomedicine (1) to large- future hold for the field of robotics? Recently, (iv) Robot swarms that allow simpler, less scale systems for robotic construction (2) and we conducted an open online survey on ma- expensive, modular units to be reconfigured from robots for outer space to those involved jor unsolved challenges in robotics. On the into a team depending on the task that needs in deep-sea exploration (3). We have seen the basis of the feedback and submissions received, to be performed while being as effective as a evolution of soft robots and how new mate- an invited online expert panel was convened, larger, task-specific, monolithic robot. rials and fabrication schemes have led to de- and the panel shortlisted the 30 most import- (v) Navigation and exploration in extreme formable actuators that are compliant, versatile, ant topics and research directions. These are environments that are not only unmapped but and self-healing­ (4–6). We have also seen many further grouped into 10 grand challenges (Fig. 1) also poorly understood, with abilities to adapt, examples of bioinspired designs, from the that may have major breakthroughs, signifi- to learn, and to recover and handle failures. Fundamental aspects of artificial in- power-modulated jumping robot with agility cant research, and/or socioeconomic impact (vi) by guest on February 1, 2018 and power that approach those of galagos (the in the next 5 to 10 years: telligence (AI) for robotics, including learn- animal with the highest vertical jumping agil- (i) New materials and fabrication schemes ing how to learn, combining advanced pattern ity) (7) to a biomimetic robotic platform to for developing a new generation of robots that recognition and model-based reasoning, and study flight specializations of bats (8) and a are multifunctional, power-efficient, compli- developing intelligence with common sense. biorobotic adhesive disc for underwater hitch­ ant, and autonomous in ways akin to biolog- (vii) Brain-computer interfaces (BCIs) for hiking­­ inspired by the remora suckerfish (9). ical organisms. seamless control of peripheral neuropros- We also celebrated the 10th anniversary of the (ii) Biohybrid and bioinspired robots that theses, functional electric stimulation devices, Robot Operating System (ROS) (10), the open-­ translate fundamental biological principles and exoskeletons. source robotics middleware that is making into engineering design rules or integrate liv- (viii) Social interaction that understands human social dynamics and moral norms and that can be truly integrated with our social life 1Hamlyn Centre for Robotic Surgery, Imperial College London, London, UK. 2Center for Marine Robotics, Woods Hole Oceanographic Institution, Woods Hole, MA 02543, USA. 3Department of Cardiovascular Surgery, Boston showing empathy and natural social behaviors. Children’s Hospital, Harvard Medical School, Boston, MA 02115, USA. 4Institute of Physical Chemistry, University (ix) Medical robotics with increasing levels of Stuttgart, Stuttgart, Germany. 5Micro, Nano, and Molecular Systems Laboratory, Max Planck Institute for of autonomy but with due consideration of legal, Intelligent Systems, Stuttgart, Germany. 6Centre for Practical Ethics, Faculty of Philosophy, University of Oxford, Oxford, UK. 7Digital Ethics Lab, Oxford Internet Institute, University of Oxford, Oxford, UK. 8Department of ethical, and technical challenges, as well as mi- Computer Science, University of Oxford, Oxford, UK. 9Data Ethics Group, Alan Turing Institute, London, UK. crorobotics tackling real demands in medicine. 10Department of Economics, American University, Washington, DC 20016, USA. 11Department of Integrative (x) Ethics and security for responsible in- 12 Biology, University of California, Berkeley, Berkeley, CA 94720, USA. Singularity University, NASA Research novation in robotics. Park, Moffett Field, CA 94035, USA. 13MediaX, Stanford University, Stanford, CA 94305, USA. 14Department of Mechanical Engineering and Applied Mechanics, University of Pennsylvania, Philadelphia, PA 19104, USA. 15Na- The field of robotics is broad and covers tional Academy of Sciences, Washington, DC 20418, USA. 16Institute of Robotics and Intelligent Systems, De- many underpinning and allied technological 17 partment of Mechanical and Process Engineering, ETH Zürich, Zurich, Switzerland. Department of Computer areas. The identification of these challenges Science, Yale University, New Haven, CT 06520, USA. 18Department Mechanical Engineering and Materials Science, Yale University, New Haven, CT 06520, USA. 19Department of Computer Science, Johns Hopkins Uni- was a difficult task, and there are many sub- versity, Baltimore, MD 21218, USA. 20Machine Learning Department, School of Computer Science, Carnegie Mel- topics not listed that are equally important lon University, Pittsburgh, PA 15213, USA. 21School of Materials Science and Engineering, Georgia Institute of to future development. The above list is there- Technology, Atlanta, GA 30332, USA. 22John A. Paulson School of Engineering and Applied Sciences, Harvard 23 fore neither exclusive nor exhaustive. University, Cambridge, MA 02138, USA. Wyss Institute for Biologically Inspired Engineering, Harvard University, Cambridge, MA 02138, USA. One of the ambitions of Science Robotics *Corresponding author. Email: [email protected] is to deeply root robotics research in science

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 1 of 14 SCIENCE ROBOTICS | PERSPECTIVE Downloaded from

Fig. 1. Ten grand challenges of Science Robotics. http://robotics.sciencemag.org/ while developing novel robotic platforms that tunable properties, and new fabrication strat- a promising material for soft and wearable will enable new scientific discoveries. Of the egies to embody these functional materials robotics, generating significant interest in em- 10 grand challenges listed here, the first seven as new capabilities for future robots, includ- bedding electrical functionality into fabrics. represent underpinning technologies that have ing the robot building and repairing itself. Finally, bidirectional transducers can enable a wider impact on all application areas of ro­ New materials that combine sensing and sensors and actuators to behave as materials botics. For the next two challenges, we have actuation challenge the physical limitations for energy harvesting or storage. While de- included social robotics and medical robotics of traditional mechatronic systems and offer veloping new materials for the future of ro- as application-specific areas of development a range of opportunities for the design of new botics, it will be important to consider the

to highlight the substantial societal and health robots (14). Many of the design principles biodegradability issues or as part of the cir- by guest on February 1, 2018 impacts that they will bring. Finally, the last draw inspiration from nature. In vertebrates, cular economy paradigm to ensure their eco-­ challenge is related to responsible innovation one finds a wide range of material properties sustainability. This is particularly relevant given and how ethics and security should be care- from soft tissue to bone—over seven orders the ubiquitous nature of robotic platforms in fully considered as we develop the technology of magnitude in modulus—that is mediated future (19, 20). further. by a continuous gradient of compliance. As Fabrication and assembly is typically a se- opposed to the more “nuts-and-bolts” assem- rial process that is slow and difficult to scale bly approaches currently used to combine basic to very large or very small scales. The 2016 CURRENT STATE OF THE ART AND components into complete robots, a seamless Nobel Prize in Chemistry was awarded to three 10 GRAND CHALLENGES integration of dissimilar material properties pioneers in the field of mechanochemistry New materials and fabrication schemes (e.g., rigid with soft, conductive with dielec- who created the first synthetic molecular ma- Gears, motors, and electromechanical actu- tric, etc.), spatially patterned with resolution chines. A major remaining challenge that has ators are fundamental to many of the robotic several orders of magnitude smaller than the thus far not been realized, despite Feynman’s platforms in use today, but laboratories around characteristic dimension of the robot, could prophecy (21), is to develop materials by inte- the world have begun to explore new materials obviate the need for complex assembly and grating these molecular machines, or other including artificial muscles (11), compliant lead to distributed function. force-generating molecules or biological motor materials for soft robots (12), and emerging Similar to functionally graded materials, proteins, into hierarchical materials. Substantial advanced manufacturing and assembly strat- multifunctional materials can increase the ef- opportunity exists in the convergence of ad- egies (13). As illustrated in Fig. 2, these promise ficiency of robot design and simultaneously ditive and subtractive methods, with emerg- a new generation of robots that are power-­ offer distributed networks of hierarchically ing technologies involving two-dimensional efficient, multifunctional, compliant, and au- structured sensors and actuators. Opportuni- (2D) to 3D transformations to generate new tonomous in ways that are similar to biological ties exist to leverage breakthroughs in folding-­ architectures that can simplify the need for organisms. However, most demonstrations based metamaterials that have demonstrated specialized hardware and enhance the robot’s using new materials and fabrication strategies tunable electromagnetic (15) or mechanical function. For example, 3D printing (or simi- have been “one-offs” and must still overcome (16) properties beyond what is possible with lar techniques such as multiphoton lithography basic hurdles to achieve wide-scale adoption. the base material itself. Similarly, multiphase or selective laser sintering) can create features These hurdles include improved portable en- composites may be used for simultaneous flu- and structures over nine orders of magnitude ergy storage and harvesting, new materials with idic actuation or sensing (17, 18). Textiles are in size. However, there is no single technique

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 2 of 14 SCIENCE ROBOTICS | PERSPECTIVE

Functionally graded Multifunctional or machine that can cover this materials density. No battery can yet match range—the best additive manu- metabolic energy generation in facturing strategy covers roughly organisms, so highly miniaturized, three or four orders of magnitude biohybrid robots actuated by bio- in scale range—and none offers logical muscle become advanta- more than a handful of mate- geous (28). Biohybrid robots can rials choices. Alternative methods exploit the unique features of liv- should be explored that combine ing cells that include self-healing techniques from micro-/nanoscale (35), embedded sensors, dynamic fabrication (e.g., surface and bulk response to changing environmen- micromachining; physical and tal conditions, and use of inexpen- chem­ical deposition; and mi­ sive and eco-friendly­ fuel (28). cro­scale molding, stamping, and A major challenge remains as functionalization used in soft li- Energy harvesting Hierarchical fabrication to how these components are ef- thography), mesoscale methods and storage fectively integrated and embodied­ such as layering and lamination to perform system-level behaviors Downloaded from common in multilayer printed (Fig. 3). The field of bioinspired circuit boards, and the myriad robotics must address different macroscale multi-axis subtractive challenges, mainly due to the syn- methods. Another challenge that thesis/fabrication of efficient and requires much more investigation scalable artificial components. is the development of multiscalar How­ever, biology has made prog- http://robotics.sciencemag.org/ materials able to adapt and heal ress toward providing principles, over time, thus providing 4D ro- Fig. 2. Multifunctional materials. New materials and fabrication schemes promise especially for mobility and manip- bots that achieve the complexity a new generation of robots that are power-efficient, multifunctional, compliant, and ulation. New discoveries in hydro-, found in natural systems (22). autonomous in ways that are similar to biological organisms. aero-, and terradynamics have led to an impressive “robo-zoo”­ of bio­ Bioinspired and biohybrid robots communication are critical and must be shared inspired robots (24, 25) benefiting from the as developed (32). Novel designs of heteroge- nonlinear, unsteady, self-stabilizing, energy As human technologies take on more neous, anisotropic, hierarchical, multifunc- storage, and return principles quantified of the characteristics of nature, nature tional materials have used differing designs of in animals. Further development is required becomes a more useful teacher (23). structural elements to increase material strength, to understand transitions and multimodal by guest on February 1, 2018 stiffness, and flexibility; fracture toughness; performance (36) within the same platform. By bioinspired robotics, we mean the use wear resistance; and energy absorption (33). Significant progress has been made in bio- of fundamental biological principles that are These advances pro­mise to provide robots with inspired, quasi-static, pick-and-place ma- translated into engineering design rules for features such as body support, weight reduc- nipulation, and grasping, but no system creating a robot that performs like a natural tion, impact pro­tection, morphological com- has integrated components sufficiently to system. If the biological understanding re- putation, and mobility.­ Techniques newly match the flexibility and dexterity of hu- sults in the direct use of biological material available to fabricate architectures at the micro-, man hands (37). to design synthetic machines, then we refer meso-, and ma­croscales include recom­binant As bioinspired robots venture beyond the to this as a biohybrid robot. Specific grand technologies, biomineralization, layer-by-layer laboratory, models of real-world, unstructured challenge lists for biorobotics have remained deposition, ori- and kirigami, self-assembly, environments will be required, but none can largely unchanged for the past 30 years—­­a bio-­templating, magnetic manipulation, freeze-­ yet adequately represent our complex, dynamic battery to match metabolic conversion,­ muscle- casting, vacuum-casting, extrusion and roll world. Although first-principle models ex- like actuators, self-healing material that manu- compaction,­­ laser engraving, additive manu- ist for hydro- and aerodynamic systems (i.e., factures itself, autonomy in any environment, facturing, actuator-embedded molding, and the Navier-Stokes equations), a similar frame- human-­like perception, and, ultimately, com- soft lithography (33). work for terradynamics (38) is required to un- putation and reasoning. For recent progress For biohybrid and bioinspired robots, ac­ derstand how bioinspired robots effectively on these and other specific challenges, we tuation and energy re­main major bottlenecks interact with the ground. Because of their refer readers to a few of the many outstand- compared with performance seen in animals staggering complexity, one of the greatest ing perspectives and reviews (4, 24–31). Here, (34). Electromagnetic­ motors are adequate challenges to extracting fundamental princi- we identify major goals that, if met, would actuators for large robots but inefficient at ples from biological systems involves model accelerate the design and implementation of small scales or in soft systems. New artifi- abstraction and dimensional reduction (39). bio­inspired and biohybrid robots at an un- cial muscles could revolutionize bioinspired Internal models can allow us to test hypoth- precedented pace. robots; current versions that have muscle-like eses and simplify control, especially when Major challenges remain for nearly all function and operate by shrink­­age or expan- placed into a dynamical systems theory frame- com­­ponent technologies (Fig. 3) that could sion of material—such as shape-­memory work (40). These models become even more enable bioinspired behavior. Materials that materials and electro-active polymers—lack important as we require simple representa-

CREDIT: A. KITTERMAN/ SCIENCE ROBOTICS CREDIT: couple sensing, actuation, computation, and robustness, efficiency, and energy and power tions for use in reinforcement, supervised,

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 3 of 14 SCIENCE ROBOTICS | PERSPECTIVE

In practice, the operational longev- ity of a mobile, autonomous system is typically dictated by the battery power, Materials Mobility: swimming, its size, and its weight. Efforts continue crawling, digging, Sensors to minimize power utilization through walking, running, Actuators s jumping, climbing, ying, development of power-e­ fficient electron- gie Sy lo st multimodal ics and actuators, but for robots to op- Fabrication o em hn c sy erate wirelessly for appreciable times in e n Manipulation Compliance t t g h unstructured environments, they must n e Energy li s i Collective and b s extract useful energy from their sur- a cooperative systems Computation n E roundings and use radical new solutions

Communication Human-robot for highly energy-dense storage, such as Bioinspired interaction solar light, vibration, and mechanical and movement. Compared with biological ma- Models of environment Biohybrid Exploration chines at any scale, robots are typically very Robots Biomedical energy-­­­inefficient [e.g., the 100-horsepower

Control systems Downloaded from (75 kW) consumption of Boston Dynam- Dynamical models Security n T o ic’s horse-sized LS3]. Whereas the quint- h ti Dimensional reduction a Biodiversity and conservation e c essential robot arm bolted to the factory or li y p Exaptation approaches Ap Environmental monitoring floor and tethered to an unlimited power supply works well in industrial settings, Decision making and creativity Structural inspection and maintenance mobile robots lack a standard fuel source, Reinforcement, supervised http://robotics.sciencemag.org/ and deep learning Social assistive and home service storage, and distribution system. Batteries, of course, are ubiquitous, although their Fig. 3. Convergence of conditions accelerating opportunities for design of bioinspired and biohybrid robots. energy density remains low compared Enabling technologies, development of theory, a synthesis of systems, and application drivers all provide a foun- with hydrocarbons (about 1 MJ/kg and dation for a frontier. [Adapted by N. Cary/Science Robotics. Image credits: Octobot (4); RoboBee, Wyss Institute at 50 MJ/kg, respectively). One benchmark Harvard University; StickyBot, PolyPEDAL Laboratory-Pauline­ Jennings, robot from M. Cutkosky, Stanford University; comes from biology, where carbohydrates 105 ray ( )]. (about 17 MJ/kg) power the effective run- ning, swimming, and flying of organisms and deep learning for adaptability, decision-­ pacities use Ni-, Li-, and Mn-rich, layered ma- over a huge range of physical scales (45).

making, and even creativity. terials (41). Although many new ideas are being Robotics will require a shift in energy stor- by guest on February 1, 2018 developed, the fundamental issues being ad- age technologies to produce similar behav- Power and energy dressed remain the same for many historical ior. Electrochemical storage technologies are As for any electronic system, power and ener- technologies: irreversible phase transitions of attractive for numerous reasons, although gy sources represent one of the most challeng- active materials and/or unstable electrode-­ many autonomous robots leverage combus- ing areas of robotics research and deployment, electrolyte solution interfaces (41). tion (13) or monopropellant decomposition especially for mobile robotics (Fig. 4). Under- Metal-oxygen, lithium-sulfur, aluminum-­ (46) as alternatives. water and particularly deep-sea exploration ion, and sodium-ion batteries are some of the Developments in energy-harvesting tech- requires compact, stable, high–energy density key technologies being actively pursued. The niques (e.g., mechanical, thermoelectric, photo­­ batteries to support robots working in chal- potential of lithium-sulfur batteries combined voltaic, and electrochemical) and wireless lenging conditions and extreme environments. with solar panels has already been demon- power transmission (47) are expected to play The increasing adoption of drones and auton- strated with the Zephyr-6 unmanned aerial a key role in addressing the power and energy omous vehicles is fueling the development of vehicle in its record-setting, high-altitude, long-­ challenges of robotics. Different mechanisms new battery technologies that are safe and af- endurance flights (42). Although most bat- have been established for harvesting mechan- fordable, with longer cycle lives, robust tem- tery research is focused on liquid electrolytes ical energy, including electromagnetic and elec- perature tolerance, higher energy densities, and because of high ion conductivity and good trostatic generators, as well as piezoelectric relatively low weight. Beyond the currently surface-wetting properties, they often suffer nanogenerators and triboelectric nanogener- available commercial technologies such as lead-­ from electrochemical and thermal instabili- ators (based on the coupled effect of contact acid, nickel–metal hydride, and lithium-ion ties, as well as low ion selectivity. Advanced electrification and electrostatic induction) (48). batteries, there has already been extensive re- battery systems based on solid electrolytes Besides serving as a small power supplies, search on developing next-generation technol- could bring advantages because of their safety, nanogenerators can be self-powered sensors ogies, such as fuel cells and supercapacitors. excellent stability, long cycle lives, and low and flexible actuators with the use of a range These new areas include the development of cost (43). The advent of flow-based, lithium-­ of materials from functional polymers, fab- silicon anodes with smart electrodes through ion, lithium-sulfur, and lithium-organic bat- rics, and nanomaterials to traditional metal conductive nanoporous structures and binder teries also promises new opportunities (44). foils and ceramic thin films (49). The most designs, which greatly enhances cyclability The future will also see new improvements to important characteristic of a nanogenerator is and minimizes pulverization. Other emerging the current radioisotope power systems used its high response to low-frequency mechanical electrode designs for achieving enhanced ca- for space exploration. triggering, with complementary applications

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 4 of 14 SCIENCE ROBOTICS | PERSPECTIVE

mainstream. Tools from data science, ma- chine learning, and predictive analytics are now being routinely used to extract information from text and speech and to recognize objects from imagery (pic- tures and videos). As we think about swarms, it is useful to consider different forms of collec- tive behavior. Coordination and co- operation can be seen in groups that are homogeneous, but heterogeneity is powerful because it allows for collabo- ration (52, 53). For example, a large ro- bot may be able to carry more powerful sensors or have more powerful compu- tational resources or radios, but it may Downloaded from be less agile than its smaller counter- parts. Scale is particularly important in robot swarms where small groups lend themselves to centralized control, and information across the group can, in prin- ciple, be shared via communication and http://robotics.sciencemag.org/ sensing. The analysis of group behav- Fig. 4. A summary of different energy sources for robotics. Power generators, which include fuel cells, classical electromagnetic generators, and solar cells. Energy storage, including batteries and capacitors/supercapacitors. Pow- iors in these settings or the synthesis of er harvesting and newly developed nanogenerators, as micro-/nano-energy sources, self-powered sensors, and flexi- group behaviors for a given task is easier ble transducers. for smaller groups with centralized ar- chitecture than for larger groups like swarms, where it is impractical to effi- with an electromagnetic generator that usu- fundamental to navigating as a flock or horde, ciently share information across the swarm and ally works well at high operation frequency. foraging, hunting, building nests, and surviv- architecture because these systems are nec- In the working environment of a robotics, ing harsh environments. Similarly, a group essarily decentralized. From a mathemati-

low-frequency­­ mechanical stimulation is of relatively unsophisticated robots can form cal perspective, the state space, which is the by guest on February 1, 2018 fairly popular, which can be effectively con- a networked team that realizes a range of Cartesian product of the individual state verted into electric output using a triboe­ lectric behaviors well beyond the capabilities of the spaces, grows linearly, and the types of in- nanogenerator. individuals by communicating and cooperat- teractions that can occur across individuals As stated in the previous section, no battery ing with team members. The swarm principle grow combinatorially. Thus, it is necessary can yet match metabolic energy generation in can be used at macro-, micro-, and nanoscales to develop stochastic models for predicting organisms. Biohybrid robots could use the with a plethora of application areas. collective behavior in large-scale swarms. unique features of living cells for potential There are three technology drivers sug- However,­ we lack mathematical models of solutions (28). gesting that robot swarms will have an im- flock- or herd-like groups that elude the enu- pact in the next 5 to 10 years that stem from meration in small-scale­ groups yet do not Robot swarms falling prices and increasing performances of justify ensemble-­averaged models seen in large-​ Robot swarms allow simpler, less expensive, sensors, processors, storage devices, and com- scale swarms. modular robotic units to be reconfigured into munications hardware. First, the integration Robotic systems are equipped with sensors a team (Fig. 5) depending on the task at hand of components for computation and storage that allow them to perceive the environment. while being as effective as a larger, task-specific, is resulting in a software-centric­ architecture They reason about the environment and monolithic robot, which may be more expen- that tightly couples computation, storage, net- take actions, forming a feedback loop that sive and have to be rebuilt depending on the working, and virtualization resources—a frame- is called a perception-action loop. Design- task. Nature provides a repertoire of examples work that is being called “hyper-convergence” ing perception-action­ loops is fundamental that illustrate this idea (50). Independently act- (51). Soon, sensors and wireless communica- to creating autonomous robots that func- ing organisms cannot achieve a goal by them- tion devices will be part of this hyper-convergence. tion in unstructured environments. Robot selves but, in coordination with other organisms, Second, we are seeing the convergence of the swarms require their communication ability can solve complex problems and complete a hardware for consumer electronics (smart to be embedded in this feedback loop. Thus, mission. This “force multiplication” requires phones, tablets, and virtual reality devices) and perception-action-communication­ loops are individuals to sense not only the environ- intelligent autonomous systems (drones, ro- key to designing multifunctional, adaptive ment but also their neighbors and to commu- bots, and self-­driving vehicles), with concur- robot swarms. There are currently no sys- nicate with other individuals in their team rent advances in 5G wireless technologies. tematic approaches for designing such mul- while acting independently. This paradigm has Third, cognitive systems relying on statis- tidimensional feedback loops across large been seen in fish, birds, and insects and is tical machine learning and AI are becoming groups.

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 5 of 14 SCIENCE ROBOTICS | PERSPECTIVE

Whether we think of smaller in tunnels or mines must cope robot groups, in which the com- with rough terrain, narrow pas- binatorics do not pose formida- sageways, and degraded percep- ble challenges, or larger swarms, tion. Robots undertaking nuclear much of the literature address- decommissioning must withstand ing the problem of coordination radiation and restricted access, makes use of simpler mathematical and those used to construct and models; algorithms for perception, assemble infrastructure must be estimation, planning, and control; able to resist chemicals and ma- and robot deployments (54). The terials used in the construction dynamics and control of cooperation process as well as being resistant to have been addressed in coopera- dirt, dust, and large impact forces. tive manipulation, multi-​fingered Undersea robots operate in an en- grasping, and legged locomotion, vironment where radio does not but systematic approaches to ques- penetrate and our usual forms of tions of synthesis do not exist. communication and navigation Downloaded from Similarly, although there is in- disappear; untethered undersea teresting work on collaboration Fig. 5. Robot swarms. New opportunities and research challenges. vehicles must be autonomous. As between humans and robots (55) robotic spacecraft take on tasks and between aerial and ground like roaming distant planetary sur- robots (56), a general framework faces and, in the not-so-distant for modeling heterogeneity and the design human civilization. They can provide solu-

future, possibly landing on the icy moons of http://robotics.sciencemag.org/ of heterogeneous groups and desired behav- tions to feed an ever-increasing population the outer planets, they enter a regime where iors does not exist. with limited resources by increasing the ef- long latency and low bandwidths of commu- As we develop robot swarms, one must also ficiency of food production and decreasing nications not only greatly reduce productiv- develop the tools to create teams that can be water consumption by an order of magni- ity but also put the survival of the robot itself responsive to human commands, can adapt tude (58). They can respond to natural di- at risk. to changing conditions, are robust to distur- sasters and adversarial attacks by enabling Undoubtedly, current mapping and nav- bances (to the extent that is possible given resilience in our infrastructure (59). They igation techniques will continue to evolve. the constraints on resources), and are resilient are a part of any practical solution to space For example, techniques such as SLAM (simul- to adversarial, disruptive changes caused by colonization. We are poised to see great ad- taneous localization and mapping) will go be-

large-scale failures or damage to the swarm vances and impacts in this area in the next 5 yond the current rigid and static assumptions by guest on February 1, 2018 infrastructure. Responsiveness is generally to 10 years. of the world and will effectively deal with time-­ characterized by the time a system takes to varying, dynamic environments with deform- respond to input or meet input-output (task) Navigation and exploration able objects (60). With resource constraints, specifications. Robustness is the property of Path planning, obstacle avoidance, localiza- specific challenges include how to learn, for- the system to be responsive even in the pres- tion, and environment mapping are ubiqui- get, and associate memories of scene content ence of disturbances and modeling errors tous requirements of robot navigation and both qualitatively and semantically, similar (and failures), although the majority of the exploration. Advances in sensing, machine to how human perception works; how to literature addresses robustness with carefully vision, and embedded computation have surpass purely geometric maps to have se- constructed bounds on those disturbances underpinned the remarkable progress of mantic understanding of the scenes; how to and modeling errors. As pointed out by Rodin autonomous vehicles roaming complex ter- reason about new concepts and their seman- (57) in the context of similar challenges that rains at speed, drones forming swarms for tic representations and discover new objects confront urban societies today, resilience is completing collaborative tasks, and surgical or classes in the environment through learn- a fundamentally different property that is robots delivering targeted therapy while ne- ing and active interactions; and how to evolve about systems that can bend without break- gotiating complex, delicate anatomical struc- through online, prospective, and lifelong con- ing. Resilient systems are self-aware and self-­ tures. Many robots we deploy are intrinsic tinuous learning. ­regulating and can recover from large-scale explorers that we send to the far reaches of For navigation, the grand challenge is to disruptions to the network. Thus, a science the planet—the deep oceans, under the Arctic handle failures and being able to adapt, learn, of resilient robot swarms must focus beyond ice pack, into volcanoes—and go where no hu- and recover (Fig. 6). For exploration, it is robust individual agents to resilient integra- man has yet tread, often under unknown and developing the innate abilities to make and tion across diverse elements of the group that extreme conditions. The associated challeng- recognize new discoveries. From a system per- leverage new mechanisms (e.g., mobility, re- es are therefore much greater than those en- spective, this requires the physical robustness configuration, sensing, communication, plat- countered today. to withstand harsh, changeable environments, form diversity, and involvement of human Foremost, the robots must operate in en- rough handl­ing, and complex manipulation. peers) for achieving macroscale resilience. vironments that are not only unmapped, but, The robots need to have significant levels of Robot networks integrated with our in- at times, their very nature is not understood. autonomy leading to complex self-monitoring, frastructure have tremendous potential for Adding to this are challenges associated with self-­reconfiguration, and repair such that there

CREDIT: JAMES LAW/UNIVERSITY OF SHEFFIELD JAMES LAW/UNIVERSITY CREDIT: solving the most pressing problems facing communications and navigation. Robots is no single point of complete failure but rather

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 6 of 14 SCIENCE ROBOTICS | PERSPECTIVE

However, we do not yet have systems that can do this easily across heterogeneous tasks and domains. AI systems that can comprehend deeply and synthesize across complex texts and narratives will prove useful in a variety of applications. We have already seen some initial examples, but the real world is both interdisciplin- ary and complex, and building robust systems of this class will prove extremely challenging. One of the enduring grand challenges in AI is to provide a coherent and com- prehensive mapping of the key mecha- nisms of human intelligence in a software system. The first key step in doing this is Downloaded from to produce a thorough account of how the neocortex actually works, including learning to learn and learning from lim- Fig. 6. Intelligent explorers. Handling failures and being able to adapt, learn, and recover are major challenges for ited examples. A recent paper on this pro- 106 navigation and exploration, especially for robots operating in extreme environments. [Reproduced from ( ) with vides some detailed and testable predictions permission]. concerning how columns in the neocor- http://robotics.sciencemag.org/ tex provide location signals that enable graceful system degradation. When possible, begin to reason about underlying interdisci- learning the structure of the world (65). We solutions need to involve control of multiple plinary mechanisms and systems dynamics. need to test theories of this type rigorously, heterogeneous robots; adaptively coordinate, Meta-learning, or learning how to learn new both in terms of neuroscience data and in interface, and use multiple assets; and share things, is a critical new AI capability not only the operation of AI software (66). In addition, information from multiple data sources of vari- with large training data sets but also with lim- much progress has been made recently in able reliability and accuracy. ited data. The challenge is to be able to learn building AI systems that understand natu- on the fly, adapting to dynamic and uncertain ral language. A key set of targets is to build AI for robotics environments. One promising approach in this systems that maintain coherent conversa-

As the underpinning technology for robotics, area has been developed based on neurosci- tions and deal with unknown environments by guest on February 1, 2018 AI is undergoing a renaissance after more ence insights about the human hippocampus and contexts. than 60 years of ongoing development. There as a predictive map of novel situations (62). Ambient intelligence and ubiquitous and is a widespread myth that AI did not work AI systems that know their own limita- networked AI and robotics (cloud robotics) for the first 50 years, but the truth is that for tions and know how to seek help could go will be critical in the development of integrated certain classes of domain- and task-specific beyond the current methods of training and heterogeneous AI and robotic services. There problems, given enough development time knowledge acquisition. These systems will know are many initial examples of cloud AIs that up- as well as computing and data resources, the how to interact, how to seek help, how to re- date situation assessments and share knowledge applications could be made to work. The ad- cover from failure, and how to become smarter. but few working examples of heterogeneous vent of deep learning methods resulted in AI systems and robots that can model their AI or robotic services that integrate smoothly remarkable levels of object recognition ac- own components and operations are critical for and reliably over time. DeepMind’s PathNet curacy (61) using hierarchical pattern recog- adaptation and evolution. We need AIs that are architecture points to systems that allow for nition that retained information coherence at able to detect their own subcomponents, model new contexts to be learned at the same time, each level of the hierarchy. The new machine-­ their operations, and modify those models if leveraging knowledge of training in other con- learning algorithms were combined with un- their structure changes. Work by Bongard et al. texts to learn much faster. precedented access to data, as well as inexpensive (63) provides an early example of this type of One of the big questions for AI is its ability and powerful computing hardware. The re- robotic system, which can discover its own to perform deep moral and social reasoning sulting progress in solving narrow classes of components and learn to use them dynamically about real-world problems. As AI and robotic AI problems has led many to think that we in locomotion. systems undergo accelerating growth in power are on the verge of solving intelligence—in all AI that can learn complex tasks on its own and capabilities, there will be an increasing its multifaceted and (still) poorly understood and with a minimum of initial training data premium on systems that can demonstrate dimensions. will prove critical for next-​generation systems. moral and social reasoning. Although human- However, we still have a long way to go Most machine-learning​ systems are data-​ in-the-loop may be a preferred design con- to replicate and exceed all the facets of intel- intensive and require massive data in order straint for systems that touch life-or-death ligence that we see in humans. Combining to learn complex tasks. DeepMind’s Alpha- situations, in autonomous driving and aero- advanced pattern recognition and model-­ Go Zero system that taught itself to play Go space applications, the relevant decision loops based reasoning is critical for building systems significantly better than the world champion may well be too fast for the human brain, that can go beyond statistical correlation and in Go (64) was an impressive example of this. hence the need for embedded moral and

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 7 of 14 SCIENCE ROBOTICS | PERSPECTIVE social reasoning. These challenges need to be ponent miniaturization, biocompatibility of Despite the success of BCI for patients framed in the context of baseline risks that materials, and embedded computing. with amyotrophic lateral sclerosis (also known humans have already habituated to, such as For practical use, a BCI can be classified as motor neuron disease), spinal cord injury, 1.2 million people dying worldwide as a result as active, reactive, or passive (73). Active BCI and rehabilitation of motor function after stroke, of largely avoidable driver errors committed derives its outputs from brain activity, which there remain significant challenges for the wid- by humans. We can expect to see consider- is directly and consciously controlled by the er adoption of BCI (74). The first is in sensing able and rapid operational progress on this user, not necessarily depending on external and data acquisition because current modal- front. events, for controlling an application. In re- ities are expensive and cumbersome. Parallel active BCI, the outputs are derived from brain developments in implantable sensing with new Brain-computer interfaces activities arising in response to specific exter- microfabrication, packaging, and flexible elec- A BCI forges a direct, online communication nal stimuli. Passive BCI is a relatively newer tronics, combined with ultralow-power local between brain and machine, independent from concept, which derives its outputs from arbi- processing and wireless data paths, would bring the user’s physical abilities, and represents a trary brain activity arising without the purpose new opportunities for completely untethered new way to augment human capabilities and of voluntary control, for enriching human- implants, providing improved patient expe- restore patient function (Fig. 7). Direct use machine interaction with implicit informa- rience and uptake in both clinical and home of brain activity to control a computer or ex- tion on the actual user state. environments. For noninvasive techniques, Downloaded from ternal device without the mediation of the pe- Both invasive and noninvasive methods newly emerging, low-cost, and ergonomically ripheral, somatomotor nervous system has are used to record brain activity. Invasive ap- designed wireless EEG and fNIRS systems major applications in enabling paralyzed pa- proaches measure the neural activities of the have shown promise for general BCI-based tients to communicate and control robotic brain by either intracortical neural interfaces robotic control. prosthetics and in rehabilitation for restoring with microelectrode arrays, which capture The second challenge is in data process- neural function (67–71). BCIs translate the user’s spike signals and local field potentials, or cor- ing and dealing with artifacts of noncerebral http://robotics.sciencemag.org/ intentions into outputs or actions by means of tical surface electrocorticography, providing origin, particularly for wearable approaches. machine-learning techniques, operating by both high temporal and spatial resolution with The data-processing challenge is also associated either presenting a stimulus to the operator and good immunity to artifacts (70). Noninvasive with the fact that cortex folding differs be- waiting for his/her response (synchronous) BCIs require no surgical implantation; typical tween individuals, as do relevant functional or continuously monitoring the operator’s cog- signals used include slow cortical potentials, maps. Furthermore, sensor locations may nitive activity and responding accordingly sensorimotor rhythms, P300 event–related po- differ across different recording sessions, (asynchronous). Beyond their clinical use, BCIs tentials, steady-state visual evoked potentials, and brain dynamics can be intrinsically non- also have emerging applications in neuroer- error-related negative evoked potentials, blood stationary. Current methods often involve ex- gonomics, communication and control, ed- oxygenation levels, and cerebral hemodynamic tended periods of training, calibration, learning,

ucation and self-regulation, as well as games changes. Common assessment methods in- and adaptation, thus making it prohibitive by guest on February 1, 2018 and entertainment (72). Despite being a rela- clude fMRI (functional magnetic resonance for general use. tively new field, recent advances in BCIs have imaging), fNIRS (functional near-infrared spec- Third, it remains to be seen whether BCI been accelerated by allied technologies, includ- troscopy), MEG (magnetoencephalography), will always outperform simpler techniques, ing neuroscience, sensor technologies and com- and EEGs (electroencephalograms) (70). such as those using eye tracking or muscle-​ based devices. The development of hy- brid BCIs may represent a viable way forward by combining with other, more mature assistive technologies. This would allow more reliable and seamless inter- facing with peripheral neuroprostheses, functional electric stimulation devices, and exoskeletons. A further challenge is dealing with tasks with high degrees of freedom. Cur- rent multiclass BCI classification gener- alizes poorly across individuals and tasks. In such cases, it may be more appropriate to rely on BCI for intention detection and task initiation and on autonomous robot manipulation for task completion. Continuing development of BCIs will bring exciting new research oppor- tunities not only in robot control and functional rehabilitation but also in knowl- Fig. 7. Brain-computer interfaces. BCIs have extensive applications in enabling paralyzed patients to communicate edge exchange and cross-fertilization with and control robotic prosthetics and in rehabilitation for restoring neural function. Continuing development of between neuroscience and robotics. BCIs will also see applications in performing mission- or safety-critical tasks. It will also play an important role in

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 8 of 14 SCIENCE ROBOTICS | PERSPECTIVE performing mission- or safety-​ Social interaction also requires critical tasks, whereby adaptive lev- building and maintaining com- els of automation, context-sens​ itive plex models of people, including decision support, and motion con- their knowledge, beliefs, goals, straints are provided depending desires, and emotions. We rou- on mental workload, task engage- tinely simplify our language based ment, hypovigilance, mood or on what we know our partners emotion, and precursors to hu- understand, coordinate our ac- man errors (e.g., hesitation and tions to match the preferences of disorientation). our collaborators, and interpret the actions of others as repre- Social interaction senting their inner goals. These Robotics and AI have often un- “hidden” states allow us not only derestimated the difficulty of rep- to understand why someone has licating capabilities that humans Fig. 8. Social robotics. Social interaction requires building and maintaining com- taken a particular action but also find particularly easy. Perhaps most plex models of people, including their knowledge, beliefs, goals, desires, and to predict their likely future be- Downloaded from notorious was the early belief that emotions. havior. Modern work on intent computer vision was a simple prob- recognition (79), user modeling lem suitable for an undergraduate in intelligent tutoring systems research project (what could be simpler than preted as a sign of uncertainty or mistrust. (80), collaboration models in human-​ seeing a table as a table?) (75), but similar stories Although we have made substantial advances machine interaction (81), emotion recognition can be told for locomotion, manipulation, and in machine perception in the last decade,

via facial feature analysis (82), and other http://robotics.sciencemag.org/ understanding language. Social interaction especially in object recognition (76), action domains touch on single aspects of this prob- has the same status: Because humans are so recognition (77), and human gaze analysis lem, but none of these domains has yet pro- adept at recognizing and interpreting social (78), we still lack systems that operate under duced comprehensive or integrated models behavior, we often underestimate the com- the diverse natural conditions and real-world that allow robots to begin to have rich, us- plexity of the challenge that this represents time constraints that social interactions de- able models of human mental states (83). for a robot (Fig. 8). mand. Next-generation systems will need to Advancing the state of the art in this do- As common as social interactions are in richly mix elements from multiple input main will require integration of models of our daily lives, we have very few comprehen- modalities and combine these perceptual sys- episodic memory, hierarchical models of sive, quantitative analyses of human social re- tems with predictive models of social inten- tasks and goals, and robust models of emo-

sponses; our understanding of human social tion to more fully capture the rich, dynamic tion to create detailed cognitive models by guest on February 1, 2018 behavior is not nearly as advanced as our nature of social interactions. that capture the naïve psychology that we knowledge of Newtonian mechanics or even Social signals are also very context-dependent effortlessly apply to understanding human human visual perception. Although this alone and culturally determined. Two individuals behavior. might make some believe that building social standing nearly nose to nose in a conversation Solutions should also work for long- interactions for robots is premature, the prac- might be typical in Argentina, but could be term interactions and relationships: A joke ticality of putting robots into our human an indication that they are either close friends told once might be funny, but the same environments—our schools, hospitals, shops, or about to have an argument in the United joke told every day for a month is not. Most and homes—necessitates addressing social States. Robots that are deployed in human en- of our current social robots have been de- interaction. The three most significant chal- vironments must be able not only to adapt signed for interaction that lasts on the order lenges that stem from building robots that to these cultural differences but also to learn of a few minutes or hours, but our human-​ interact socially with people are modeling the appropriate social and moral norms for human social interactions often span months, social dynamics, learning social and mor- their setting. A robot that expresses excite- years, and even decades. Just as machine al norms, and building a robotic theory of ment when the death of a family member is learning struggles to scale to continuous, mind. being discussed, one that shouts at inappro- long-term adaptation models (84), social Social interaction is a major challenge priate times, or one that takes a coffee mug robotics must expand from moment-to-​ for robotics in part because the perceptual de- before it is empty will not find itself wel- moment engagements to long-term relation- mands are so significant. Social cues—such come in home or workplace. The develop- ships. This expansion will require models as gaze direction, facial expressions, or vocal ment of robots capable of understanding of robot behavior and personality that dis- intonation—are often extremely detailed, empathy, ownership, and the need to keep a tinguish between changes that are appro- rapid, and nuanced signals that are embed- promise will be essential to building the long- priate at different time scales, the capability ded within other activity; the difference be- term trust and relationships necessary for op- to autonomously generate interaction con- tween an enthusiastic greeting and a sarcastic erating side by side with people. To take the tent (both verbal and nonverbal) rather than scolding might depend on a single wink, next step in this domain, new tools are re- relying on prescripted components, and per- or rising inflection on just one phoneme. quired for modeling the expectations of the sonalization and adaptation mechanisms The temporal patterning of these signals is people around the robot and expanding the that adjust both short-term responses and also frequently significant—a small delay robot’s understanding of the consequences of long-term tendencies based on current when answering a question may be inter- its own actions. interactions.

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 9 of 14 SCIENCE ROBOTICS | PERSPECTIVE

Medical robotics automation), teleoperation has been replaced Although medical robot autonomy is of- by semiautonomous or autonomous opera- ten discussed within the context of surgery, From minimally invasive surgery, tar- tion. Autonomy in medical robotics is incred- emergency medicine provides another set of geted therapy, hospital optimization, ibly challenging (88); whereas products and challenges and opportunities. In this case, an to emergency response, prosthetics, and assembly lines can be designed to fit the ca- emergency medical technician (EMT) needs home assistance, medical robotics rep- pabilities of robots, this is not possible with to assess the condition of a patient quickly, resents one of the fastest growing sec- the human body. Consequently, autonomy in prioritize problems, and often take time- tors in the medical devices industry (85). existing medical robots remains limited. In urgent steps to stabilize the patient. Intelligent most cases, the contribution of the robot has robotic systems that could assist with such The impact of robotics on medicine is unde- been to enhance the skill level of the surgeon. tasks as placing and monitoring sensors, in- niable. The therapeutic and commercial suc- For example, Intuitive Surgical’s da Vinci ro- serting intravenous lines or breathing tubes, cess of Intuitive Surgical’s da Vinci system has bot makes laparoscopy easy (89); routine pro- and preparing a patient for transport could spurred a number of commercial ventures cedures can be performed at a higher level of significantly improve the ability of an EMT targeting surgical applications, which echo the proficiency, and difficult cases that would to provide urgent care. In addition to obvious emerging trend in precision surgery, focusing otherwise be treated with open surgery can be challenges in dexterity and device develop- on early malignancies with minimally inva- performed laparoscopically. Similarly, Stryker’s ment, there are also difficult computational Downloaded from sive intervention and greater consideration Mako robotic arm enhances hip and knee challenges. The robot assistant will need to of patient recovery and quality of life (86, 87). replacement by enabling more precise bone recognize relevant patient anatomy in what These efforts will continue to improve health- drilling than the surgeon can perform on his is often a highly unstructured environment. care in terms of both outcomes and cost. Other or her own. In both these examples, the robot It will need to use its situational understand- research and commercial efforts are focusing acts as an extension of the surgeon’s hand, ing to perform tasks appropriately under on what many see as an inevitable future in and its motion is continuously under the direction of the EMT, who is likely to rely pri- http://robotics.sciencemag.org/ which intelligent robotic devices assist health- surgeon’s control. Other systems, such as Think marily on spoken commands, supplement- care workers in a variety of ways. As we move Surgical’s Robodoc system, execute precom- ed with gestures, to explain what needs to toward this future, however, many grand chal- puted and surgeon-approved cutting paths be done. lenges remain. One of the primary challenges based on medical images. All these systems A long-term challenge is to enable one in surgical and interventional robotics is a exercise some degree of “autonomy” in trans- surgeon to supervise a set of robots that can move toward systems that exhibit increas- lating a surgeon’s intentions (expressed in perform routine procedure steps autonomous- ingly higher degrees of autonomy (85). A joystick motions or in preoperative planning) ly and only call on the surgeon to take con- second grand challenge is the creation of ful- into the actual motions of the robot’s actua- trol during critical, patient-specific steps. For ly implantable robots that replace, restore, tors. The challenge arises when the controller example, intracardiac interventions involve

or enhance physiological processes. A third needs to make more complex decisions in in- navigating from percutaneous entry in the by guest on February 1, 2018 grand challenge is in the realization of micro- terpreting the clinician’s intentions. Thus, we peripheral vasculature to specific locations and nanorobotic devices of clinical relevance anticipate that the development of autonomy inside the heart using a combination of pre- (Fig. 9). in medicine satisfying regulatory and ethical and intraoperative imaging. The theory of In those industries in which robots are most concerns will progress in stages. Two exam- image-based robot navigation is well devel- successful (e.g., manufacturing and warehouse ples are described below. oped, so developing safe navigation algorithms seems quite feasible. As clinical experi- ence with intracardiac devices (e.g., tran- scatheter valves) grows, the deployment of these devices may become sufficiently standardized to enable automated de- ployment. Furthermore, miniaturized and multifunctional fully implantable robots represent an emerging area of develop- ment (90, 91). Issues related to biocom- patibility, packaging, power efficiency, and harvesting are important to be ad- dressed (92). Perhaps the most significant challenge of automating any clinical task is to be able to anticipate, detect, and re- spond to all possible failure modes. Med- ical device regulation of autonomous robots will likely need to develop in a man- ner that balances the requirements for provably safe algorithms with compli- ance costs. Fig. 9. Medical robotics. From macro to micro and from large systems to small, smarter devices that can support the An emerging area of medical robot- future development of precision medicine. ics is implantable robotic devices. These

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 10 of 14 SCIENCE ROBOTICS | PERSPECTIVE

bionic systems are being proposed as re- thus lowering inequality, whereas social placement organs, e.g., for the pancreas (91); Removing solidarity should ensure that AI’s costs as assist devices for damaged organs, e.g., for human Devaluing are shouldered by future generations, the heart (90); and to induce organ growth, responsibility human too, because they will profit enormously e.g., of the esophagus and bowel (93). There skills from it. are a number of challenges that must be ad- Reducing Fourth, AI may erode human free- dressed to advance this field. These include human 5 Ethical dom, because it may lead to unplanned control biocompatibility, reliability, adaptability, se- Challenges Eroding and unwelcome changes in human be- curity, and providing power. Full biocom- of Articial human self- haviors to accommodate the routines patibility is important in order to maintain Intelligence determination that make automation work and peo- long-term functionality. Furthermore, for ple’s lives easier. AI’s predictive power those implants that provide temporary phys- Enabling and relentless nudging, even if uninten- iological support, designing the implant to be human + tional, should foster and not undermine wrongdoing resorbable could eliminate the need for sur- Security human dignity and self-determination. gery to remove the device. Implants must Problems Finally, there is straightforward mis- also be designed to react to changing condi- use. Strictly speaking, this is not a prob- Escalation Lack of Downloaded from tions, such as exercise, and extreme reliabil- control lem with AI’s smart agency, but with the ity is a necessity because malfunction could unethical application of AI by those who quickly lead to death. Although remote pro- control it. The issues under this head- gramming to provide software updates is ad- Fig. 10. Ethical and security risks of robotics and AI ing refer to “the human use of human vantageous, security is critically important to developments. beings,” to cite the title of Wiener’s far- prevent one’s organ from being hacked. Last,

sighted book (101). Examples range from http://robotics.sciencemag.org/ because the power requirements of a robotic scanning citizens’ faces in illiberal re- device are high in comparison to, e.g., a pace- ness, ownership, privacy, surveillance, and trust gimes to discriminating among applicants maker, the capability for wireless power trans- (98). In terms of ethics, robotics and AI pose for a job or punishing law offenders unfairly. fer will be crucial. five increasingly pressing topics (Fig. 10). In this case, Kant provides the right antidote: An other emerging area of medical ro- First, excessive reliance on robotics and AI AI should be designed and used to treat every botics is micro- and nanorobotics, with in- may lead to the delegation of sensitive tasks human being always as an end and never only creasing numbers of groups maintaining to autonomous systems that should remain as a means. high-­profile research efforts. The field has at least partly subject to human supervision, In terms of security, AI can improve se- made impressive strides over the past decade either “in the loop” for monitoring purposes or curity by increasing systems’ resilience (endur-

as researchers have created a variety of small “post-loop” for redressing. Thus, it is problem- ing attacks) and robustness (averting attacks) by guest on February 1, 2018 devices capable of locomotion within liquid atic that the European Union (EU) General and combining both with counterthreat strat- environments (94). Robust fabrication tech- Data Protection Regulation does not include egies. Thanks to its autonomy, fast-paced threat niques have been developed, some devices an explicit right to an explanation when deci- analysis, and decision-making capabilities, have been functionalized for potential ap- sions affecting people are reached “solely” al- AI can enable systems verification and patch- plications (95), and therapies are being ac- gorithmically (99). ing and counter incoming threats by exploit- tively considered (96). Although excitement Second, robotics and AI may de-responsibilize ing the vulnerabilities of antagonist systems. remains high for this field, it faces a number people whenever an autonomous system could However, two challenges may hamper AI’s of significant challenges that must be addressed be blamed for a failure. A recent EU proposal potential for security. One is escalation: Ro- head-on to make continued progress toward to treat forms of AI as “electronic persons” botics and AI can refine strategies and launch clinical relevance. The primary roadblocks would only exacerbate this problem. Instead, more aggressive counteroperations. This may to overcome include the development of bio- new forms of distributed responsibility need snowball into an intensification of attacks and degradable and noncytotoxic microrobots, to be developed, learning from the legal anal- responses, which, in turn, may threaten key development of autonomous devices capable ysis of strict liability (100). infrastructures of our societies (102). The of self-directed targeting, catheter-based de- Third, unemployment is an ethical prob- solution may be to use AI to strengthen de- livery of microrobots near the target, track- lem, not just an economic one. Robotics and terring strategies and discourage opponents ing and control of swarms of devices in vivo, AI could change the workforce structure, cause before they attack, rather than mitigating the and the pursuit of clinically relevant therapies. a shift in the skills base, and potentially facil- consequences of successful attacks afterward. itate a complete de-skilling of the work force The other challenge is lack of control. Perva- Robot ethics and security even in safety-critical contexts; however, this sive distribution, multiple interactions, and With increasing levels of autonomy and human- would be imprudent. Radiologists need to keep fast-paced execution will make control of AI robot interaction, there needs to be careful con- studying images for the same reason pilots systems progressively less effective while in- sideration of potential regulatory, ethical, and need to keep landing airplanes so that they creasing the risks for unforeseen consequences legal barriers and the context of how robots are still can even if the AI cannot, or if the AI gets and errors. Regulations may mitigate the deployed. Because robotics and AI are fueled it wrong. According to a recent report, AI could lack of control by ensuring proportionality by data, some challenges are rooted in human- displace between 400 and 800 million jobs. of responses, the legitimacy of targets, and environment interactions and data governance Fairness dictates sharing the economic ben- a higher degree of responsible behavior, but

SCIENCE ROBOTICS N. CARY/ CREDIT: (97), especially consent, discrimination, fair- efits of this huge and rapid transformation, it is crucial to start shaping and enforcing

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 11 of 14 SCIENCE ROBOTICS | PERSPECTIVE policies and norms for the use of AI in secu- icry in the absence of guiding principles is dis- telligence to Benefit People and Society are rity as soon as possible while the technology couraged. Breathtaking progress is being made working on the ethics of robotics and AI. is still nascent. on relevant grand challenges in organismal bi- As with any technological innovation, the ology, but much remains unknown given the advantages of robotics and AI enable us to complexity and constraints. Biologists should not do (or do less of) things that we do not DISCUSSION AND CONCLUSIONS not only share these advances but also reveal want to do, like driving a car, and to do (either The general field of robotics is quickly evolv- how direct, comparative, and phylogenetic ex- at all or better) things that we want to do, like ing, which makes the identification of key periments using biodiversity are used to ex- enjoying a safe and secure life. In both cases, grand challenges particularly difficult. In this tract a principle. Particularly important for robotics and AI can help us tackle the many article, we have focused mainly on underpin- robotics is the development of a synergy where concrete evils oppressing humanity and our ning technologies that may have wider im- biological principles inspire novel robot or com- planet, from environmental disasters to finan- pacts across different application domains in ponent design, and these robots (or their parts) cial crises and from crime, terrorism, and war the next 5 to 10 years. There are, of course, are then used by biologists as physical mod- to famine, poverty, ignorance, inequality, and many domain-specific robotics challenges that els to further test hypotheses of biological appalling living standards. Robotics and AI need to be addressed, such as those related to structure-function relationships. This real- can and will help us manage the increasing space and marine sciences, digital architec- ization in biology—that bioinspired robots complexity of our societies, from megacities Downloaded from ture and construction, humanoids, human as- are invaluable physical models for pursuing to industrial production. Yet, the risk remains sistance, rehabilitation, agrifood, infrastructure, further advances in understanding structure-​ that we may misuse or underuse robotics and and robots designed for emergency response function, ecology, neuroethology, etc.—is also AI. We should be worried about real human and disaster relief. However, truly address- found in physics: The term “robophysics” first ignorance, not fanciful artificial superintel- ing these grand challenges requires scientists emerged (103) for the use of robots as tools to ligence. Churchill once said that “we shape and researchers from many disciplines to form study concepts in the terramechanics of lo-

our buildings and afterwards our buildings http://robotics.sciencemag.org/ ongoing collaborations. comotion, particularly on complex granular shape us” (104); this applies to robotics and When Scott, the legendary polar explorer, media. AI as well. We must design and use robotics died of exhaustion in the Antarctic, he and If bioinspired and biohybrid robots are to and AI ethically and securely and do so now. his team were within sight of their supply tent. move beyond proofs of concept and one-off Humans, not technology, are both problem Their ponies had died early in the expedition, laboratory demonstrations into real-world ap- and solution and shall remain so for any fore- and his team had to pull their heavy sleds plications, then we must match robot capa- seeable future. across the frozen landscape acting as human bility with need while not compromising pack animals. What did they carry that was so curiosity-based research. Bioinspired and bio- REFERENCES AND NOTES important it could not be left behind? Buried hybrid robots will be uniquely situated for 1. J. Li, B. E.-F. de Ávila, W. Gao, L. Zhang, J. Wang, Micro/ nanorobots for biomedicine: Delivery, surgery, under the canvas of their sled were rocks exploration, environmental monitoring, bio- by guest on February 1, 2018 containing fossils of leaves, showing that the diversity conservation, structural inspection sensing, and detoxification. Sci. Robot. 2, eaam6431 barren Antarctic continent had at one time and maintenance, security, social assistance (2017). 2. S. J. Keating, J. C. Leland, L. Cai, N. Oxman, Toward been much warmer and had once had forests. and home service, and a wide range of bio- site-specific and self-sufficient robotic fabrication on Although Scott and his team lost the race to medical applications. Market estimates fore- architectural scales. Sci. Robot. 2, eaam8986 (2017). be first to the South Pole, they made one of cast that bioinspired designs could account 3. N. Jacobstein, J. Bellingham, G.-Z. Yang, Robotics for the great discoveries of Antarctic exploration. for a substantial part of U.S. and global gross space and marine sciences. Sci. Robot. 2, eaan5594 (2017). What is notable, besides their determination domestic product (GDP) and result in mil- 4. C. Laschi, B. Mazzolai, M. Cianchetti, Soft robotics: to bring back the fossils, is that they recog- lions of future jobs. If we can meet the grand Technologies and systems pushing the boundaries of nized their significance. Such is the spirit we challenge of developing bioinspired and bio- robot abilities. Sci. Robot. 1, eaah3690 (2016). should bear in mind while pursuing these hybrid robots—and if we can establish a strong 5. E. W. Hawkes, L. H. Blumenschein, J. D. Greer, challenges: The ability to recognize discover- partnership between basic research in bio- A. M. Okamura, A soft robot that navigates its environment through growth. Sci. Robot. 2, eaan3028 (2017). ies as we progress is as important as conquer- inspired engineering and industry—then the 6. S. Terryn, J. Brancart, D. Lefeber, G. Van Assche, ing these academic missions. impact will be felt far beyond consumers and B. Vanderborght, Self-healing soft pneumatic robots. Addressing these grand challenges also re- affect many areas of engineering, science, and Sci. Robot. 2, eaan4268 (2017). quires a major cultural shift. For example, to social science as our human and natural tech- 7. D. W. Haldane, M. M. Plecnik, J. K. Yim, R. S. Fearing, Robotic vertical jumping agility via series-elastic power meet the challenges of bioinspired and bio- nologies converge. modulation. Sci. Robot. 1, eaag2048 (2016). hybrid robot design, engineers, physicists, ap- In this article, we have also highlighted the 8. A. Ramezani, S.-J. Chung, S. Hutchinson, A biomimetic plied mathematicians, and biologists must form importance of robot ethics and security. Given robotic platform to study flight specializations of bats. mutually beneficial interdisciplinary collab- the rapid pace of development in robotics Sci. Robot. 2, eaal2505 (2017). orations. To extract principles, understand a and general public concerns, it is timely that 9. Y. Wang, X. Yang, Y. Chen, D. K. Wainwright, C. P. Kenaley, Z. Gong, Z. Liu, H. Liu, J. Guan, T. Wang, J. C. Weaver, biological design, and use biological material this challenge is addressed in synchrony by R. J. Wood, L. Wen, A biorobotic adhesive disc for effectively, it is first necessary to understand basic science researchers, engineers, legal pro- underwater hitchhiking inspired by the remora that evolution is not engineering. Evolution fessionals and policy makers. Initiatives like suckerfish. Sci. Robot. 2, eaan8072 (2017). works on the principle of sufficiency, not op- AI4People, the IEEE (Institute of Electrical 10. L. Zhang, R. Merrifield, A. Deguet, G.-Z. Yang, Powering the world’s robots—10 years of ROS. Sci. Robot. 2, timality, and organisms are severely constrained and Electronics Engineers) Global Initiative eaar1868 (2017). by their complex histories, development, and on Ethics of Autonomous and Intelligent Sys- 11. T. Mirfakhrai, J. D. W. Madden, R. H. Baughman, Polymer multifunctionality. Therefore, engineered mim- tems, and the Partnership on Artificial In- artificial muscles. Mater. Today 10, 30–38 (2007).

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 12 of 14 SCIENCE ROBOTICS | PERSPECTIVE

12. C. Majidi, Soft robotics: A perspective—Current trends 33. S. E. Naleway, M. M. Porter, J. McKittrick, M. A. Meyers, 56. A. Prorok, M. A. Hsieh, V. Kumar, The impact of diversity and prospects for the future. Soft Robot. 1, 5–11 (2013). Structural design elements in biological materials: on optimal control policies for heterogeneous robot 13. N. W. Bartlett, M. T. Tolley, J. T. B. Overvelde, J. C. Weaver, Application to bioinspiration. Adv. Mater. 27, 5455–5476 swarms. IEEE Trans. Robot. 33, 346–358 (2017). B. Mosadegh, K. Bertoldi, G. M. Whitesides, R. J. Wood, (2015). 57. J. Rodin, The Resilience Dividend: Being Strong in a World A 3D-printed, functionally graded soft robot powered by 34. M. H. Dickinson, C. T. Farley, R. J. Full, M. A. R. Koehl, Where Things Go Wrong (Public Affairs, 2014). combustion. Science 349, 161–165 (2015). R. Kram, S. Lehman, How animals move: An integrative 58. T. Kozai, G. Niu, M. Takagaki, Plant Factory: An Indoor 14. G.-Z. Yang, P. Fischer, B. Nelson, New materials for view. Science 288, 100–106 (2000). Vertical Farming System for Efficient Quality Food next-generation robots. Sci. Robot. 2, eaap9294 (2017). 35. D. L. Taylor, Self‐healing hydrogels. Adv. Mater. 28, Production (Academic Press, 2015). 15. S. Yao, X. Liu, S. V. Georgakopoulos, M. M. Tentzeris, A 9060–9093 (2016). 59. Q. Zhu, T. Başar, Robust and resilient control design for novel tunable origami accordion antenna, Proceedings 36. K. H. Low, T. Hu, S. Mohammed, J. Tangorra, M. Kovac, cyber-physical systems with an application to power of the 2014 IEEE Antennas and Propagation Society Perspectives on biologically inspired hybrid and systems, 2011 50th IEEE Conference on Decision and International Symposium (APSURSI), Memphis, TN, 6 to multi-modal locomotion. Bioinspir. Biomim. 10, 020301 Control and European Control Conference (CDC-ECC), 11 July 2014 (IEEE, 2014). (2015). Orlando, FL, 12 to 15 December 2011 (IEEE, 2012). 16. J. L. Silverberg, A. A. Evans, L. McLeod, R. C. Hayward, 37. F. J. Valero-Cuevas, M. Santello, On neuromechanical 60. C. Cadena, L. Carlone, H. Carrillo, Y. Latif, D. Scaramuzza, T. Hull, C. D. Santangelo, I. Cohen, Using origami design approaches for the study of biological and robotic grasp J. Neira, I. Reid, J. J. Leonard, Past, present, and future of principles to fold reprogrammable mechanical and manipulation. J. Neuroeng. Rehabil. 14, 101 (2017). simultaneous localization and mapping: Toward the metamaterials. Science 345, 647–650 (2014). 38. C. Li, T. Zhang, D. I. Goldman, A terradynamics of legged robust-perception age. IEEE Trans. Robot. 32, 1309–1332 17. P. Polygerinos, N. Correll, S. A. Morin, B. Mosadegh, locomotion on granular media. Science 339, 1408–1412 (2016). C. D. Onal, K. Petersen, M. Cianchetti, M. T. Tolley, (2013). 61. A. Krizhevsky, I. Sutskever, G. Hinton, Imagenet

R. F. Shepherd, Soft robotics: Review of fluid‐driven 39. R. J. Full, D. E. Koditschek, Templates and anchors: classification with deep convolutional neural networks, Downloaded from intrinsically soft devices; manufacturing, sensing, Neuromechanical hypotheses of legged locomotion on in Advances in Neural Information Processing Systems 25, control, and applications in human‐robot interaction. land. J. Exp. Biol. 202, 3325–3332 (1999). P. Bartlett, K. Q. Weinberger, C. J. C. Burges, L. Bottou, Adv. Eng. Mater. 19, 1700016 (2017). 40. P. Holmes, R. J. Full, D. Koditschek, J. Guckenheimer, The F. C. N. Pereira, Eds. (Neural Information Processing 18. H.-W. Huang, M. S. Sakar, A. J. Petruska, S. Pané, dynamics of legged locomotion: Models, analyses, and Systems, 2012), pp. 1097–1105. B. J. Nelson, Soft micromachines with programmable challenges. SIAM Rev. 48, 207–304 (2006). 62. K. L. Stachenfeld, M. M. Botvinick, S. J. Gershman, The motility and morphology. Nat. Commun. 7, 12263 41. J. W. Choi, D. Aurbach, Promise and reality of hippocampus as a predictive map. Nat. Neurosci. 20, (2016).

post-lithium-ion batteries with high energy densities. 1643–1653 (2017). http://robotics.sciencemag.org/ 19. J. Rossiter, J. Winfield, I. Ieropoulos, Here today, gone Nat. Rev. Mater. 1, 16013 (2016). 63. J. Bongard, V. Zykov, H. Lipson, Resilient machines tomorrow: Biodegradable soft robots. Proc. SPIE 9798, 42. X. Zhu, Z. Guo, Z. Hou, Solar-powered airplanes: A through continuous self-modeling. Science 314, 97981S (2016). historical perspective and future challenges. Prog. 1118–1121 (2006). 20. M. Irimia-Vladu, “Green” electronics: Biodegradable and Aerospace Sci. 71, 36–53 (2014). 64. D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, biocompatible materials and devices for sustainable 43. A. Manthiram, X. Yu, S. Wang, Lithium battery A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton, future. Chem. Soc. Rev. 43, 588–610 (2014). chemistries enabled by solid-state electrolytes. Nat. Rev. Y. Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche, 21. R. Feynman, Infinitesimal machinery. J. Microelectromech. Mater. 2, 16103 (2017). T. Graepel, D. Hassabis, Mastering the game of Go Syst. 2, 4–14 (1993). 44. M. Park, J. Ryu, W. Wang, J. Cho, Material design and without human knowledge. Nature 550, 354–359 22. A. S. Gladman, E. A. Matsumoto, R. G. Nuzzo, engineering of next-generation flow-battery (2017). L. Mahadevan, J. A. Lewis, Biomimetic 4D printing. Nat. technologies. Nat. Rev. Mater. 2, 16080 (2016). 65. J. Hawkins, S. Ahmad, Y. Cui, A theory of how columns Mater. 15, 413–418 (2016). 45. V. A. Tucker, The energetic cost of moving about: in the neocortex enable learning the structure of the 23. S. Vogel, Cats’ Paws and Catapults: Mechanical Worlds of Walking and running are extremely inefficient forms world. Front. Neural Circuits 11, 81 (2017). by guest on February 1, 2018 Nature and People (WW Norton & Company, 1998). of locomotion. Much greater efficiency is achieved 66. B. M. Lake, T. D. Ullman, J. B. Tenenbaum, 24. R. Pfeifer, M. Lungarella, F. Iida, Self-organization, by birds, fish—and bicyclists. Am. Sci. 63, 413–419 S. J. Gershman, Building machines that learn and think embodiment, and biologically inspired robotics. Science (1975). like people. Behav. Brain Sci. 40, e253 (2017). 318, 1088–1093 (2007). 46. M. Wehner, R. L. Truby, D. J. Fitzgerald, B. Mosadegh, 67. J. P. Donoghue, Connecting cortex to machines: Recent 25. A. J. Ijspeert, Biorobotics: Using robots to emulate and G. M. Whitesides, J. A. Lewis, R. J. Wood, An integrated advances in brain interfaces. Nat. Neurosci. 5, investigate agile locomotion. Science 346, 196–203 design and fabrication strategy for entirely soft, 1085–1088 (2002). (2014). autonomous robots. Nature 536, 451–455 (2016). 68. M. D. Serruya, N. G. Hatsopoulos, L. Paninski, 26. S. Kim, C. Laschi, B. Trimmer, Soft robotics: A bioinspired 47. M. Boyvat, J.-S. Koh, R. J. Wood, Addressable wireless M. R. Fellows, J. P. Donoghue, Brain-machine interface: evolution in robotics. Trends Biotechnol. 31, 287–294 actuation for multijoint folding robots and devices. Instant neural control of a movement signal. Nature (2013). Sci. Robot. 2, eaan1544 (2017). 416, 141–142 (2002). 27. D. Rus, M. T. Tolley, Design, fabrication and control of 48. Z. L. Wang, Triboelectric nanogenerators as new energy 69. M. Velliste, S. Perel, M. C. Spalding, A. S. Whitford, soft robots. Nature 521, 467–475 (2015). technology for self-powered systems and as active A. B. Schwartz, Cortical control of a prosthetic arm for 28. L. Ricotti, B. Trimmer, A. W. Feinberg, R. Raman, mechanical and chemical sensors. ACS Nano 7, self-feeding. Nature 453, 1098–1101 (2008). K. K. Parker, R. Bashir, M. Sitti, S. Martel, P. Dario, 9533–9557 (2013). 70. U. Chaudhary, N. Birbaumer, A. Ramos-Murguialday, A. Menciassi, Biohybrid actuators for robotics: A review 49. Z. L. Wang, L. Lin, J. Chen, S. Niu, Y. Zi, Triboelectric Brain–computer interfaces for communication and of devices actuated by living cells. Sci. Robot. 2, Nanogenerators (Springer, 2016). rehabilitation. Nat. Rev. Neurol. 12, 513–525 (2016). eaaq0495 (2017). 50. J. K. Parrish, L. Edelstein-Keshet, Complexity, pattern, 71. J. N. Mak, J. R. Wolpaw, Clinical applications of 29. Y. Mengüç, N. Correll, R. Kramer, J. Paik, Will robots be and evolutionary trade-offs in animal aggregation. brain-computer interfaces: Current state and future bodies with brains or brains with bodies? Sci. Robot. 2, Science 284, 99–101 (1999). prospects. IEEE Rev. Biomed. Eng. 2, 187–199 (2009). eaar4527 (2017). 51. https://en.wikipedia.org/wiki/Hyper-converged_ 72. L. F. Nicolas-Alonso, J. Gomez-Gil, Brain computer 30. R. Raman, R. Bashir, Biomimicry, biofabrication, and infrastructure. interfaces, a review. Sensors 12, 1211–1279 (2012). biohybrid systems: The emergence and evolution of 52. J. P. Desai, J. P. Ostrowski, V. Kumar, Modeling and 73. T. O. Zander, C. Kothe, S. Jatzev, M. Gaertner, Enhancing biological design. Adv. Healthc. Mater. 6, 1700496 control of formations of nonholonomic mobile robots. human-computer interaction with input from active (2017). IEEE Trans. Rob. Autom. 17, 905–908 (2001). and passive brain-computer interfaces, in Brain- 31. S. Kim, M. Spenko, S. Trujillo, B. Heyneman, V. Mattoli, 53. A. Jadbabaie, J. Lin, A. S. Morse, Coordination of groups Computer Interfaces, D. S. Tan, A. Nijholt, Eds., (Springer, M. R. Cutkosky, Whole body adhesion: Hierarchical, of mobile autonomous agents using nearest neighbor 2010), pp. 181–199. directional and distributed control of adhesive forces rules. IEEE Trans. Automat. Contr. 48, 988–1001 (2003). 74. E. Vaadia, N. Birbaumer, Grand challenges of brain for a climbing robot, 2007 IEEE International Conference 54. S. Berman, Q. Lindsey, M. S. Sakar, V. Kumar, S. C. Pratt, computer interfaces in the years to come. Front. on Robotics and Automation, Rome, Italy, 10 to 14 April Experimental study and modeling of group retrieval in Neurosci. 3, 151–154 (2009). 2007 (IEEE, 2007). ants as an approach to collective transport in swarm 75. S. A. Papert, The summer vision project (1966); http:// 32. M. A. McEvoy, N. Correll, Materials that couple sensing, robotic systems. Proc. IEEE 99, 1470–1481 (2011). hdl.handle.net/1721.1/6125 actuation, computation, and communication. Science 55. A. Bauer, D. Wollherr, M. Buss, Human–robot collaboration: 76. P. Loncomilla, J. Ruiz-del-Solar, L. Martínez, Object 347, 1261689 (2015). A survey. Int. J. Human. Robot. 5, 47–66 (2008). recognition using local invariant features for robotic

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 13 of 14 SCIENCE ROBOTICS | PERSPECTIVE

applications: A survey. Pattern Recogn. 60, 499–514 robotic soft tissue surgery. Sci. Transl. Med. 8, 337ra64 99. S. Wachter, B. Mittelstadt, L. Floridi, Transparent, (2016). (2016). explainable, and accountable AI for robotics. Sci. Robot. 77. S. Herath, M. Harandi, F. Porikli, Going deeper into 89. K. P. Sajadi, H. B. Goldman, Robotic pelvic organ 2, eaan6080 (2017). action recognition: A survey. Image Vis. Comput. 60, prolapse surgery. Nat. Rev. Urol. 12, 216–224 100. L. Floridi, Distributed morality in an information society. 4–21 (2017). (2015). Sci. Eng. Ethics 19, 727–743 (2013). 78. H. Admoni, B. Scassellati, Social eye gaze in 90. C. J. Payne, I. Wamala, D. Bautista-Salinas, M. Saeed, 101. N. Wiener, The Human Use of Human Beings (Houghton human-robot interaction: A review. J. Hum. Robot D. Van Story, T. Thalhofer, M. A. Horvath, C. Abah, Mifflin, 1954). Interact. 6, 25–63 (2017). P. J. del Nido, C. J. Walsh, N. V. Vasilyev, Soft robotic 102. M. Taddeo, The limits of deterrence theory in 79. G. Sukthankar, C. Geib, H. H. Bui, D. Pynadath, ventricular assist device with septal bracing for therapy cyberspace. Philos. Technol. 2017, 1–17 (2017). R. P. Goldman, Plan, Activity, and Intent Recognition: of heart failure. Sci. Robot. 2, eaan6736 (2017). 103. J. Aguilar, T. Zhang, F. Qian, M. Kingsbury, B. McInroe, Theory and Practice (Elsevier, 2014). 91. V. Iacovacci, L. Ricotti, P. Dario, A. Menciassi, Design and N. Mazouchova, C. Li, R. Maladen, C. Gong, M. Travers, 80. S. Rossi, F. Ferland, A. Tapus, User profiling and development of a mechatronic system for noninvasive R. L. Hatton, A review on locomotion robophysics: behavioral adaptation for HRI: A survey. Pattern refilling of implantable artificial pancreas. IEEE ASME The study of movement at the intersection of robotics, Recognit. Lett. 99, 3–12 (2017). Trans. Mechatron. 20, 1160–1169 (2015). soft matter and dynamical systems. Rep. Prog. Phys. 79, 81. B. Hayes, B. Scassellati, Autonomously constructing 92. G.-Z. Yang, Implantable Sensors and Systems—From 110001 (2016). hierarchical task networks for planning and Theory to Practice (Springer, 2018). 104. House of Commons Rebuilding, http://hansard. human-robot collaboration, 2016 IEEE International 93. D. D. Damian, K. Price, S. Arabagi, I. Berra, Z. Machaidze, millbanksystems.com/commons/1943/oct/28/ Conference on Robotics and Automation (ICRA), S. Manjila, S. Shimada, A. Fabozzo, G. Arnal, D. Van Story, house-of-commons-rebuilding. Stockholm, Sweden, 16 to 21 May 2016 (IEEE, 2016). J. D. Goldsmith, A. T. Agoston, C. Kim, R. Jennings, 105. S.-J. Park, M. Gazzola, K. S. Park, S. Park, V. Di Santo,

82. C. A. Corneanu, M. O. Simón, J. F. Cohn, S. E. Guerrero, P. D. Ngo, M. Manfredi, P. E. Dupont, In vivo tissue E. L. Blevins, J. U. Lind, P. H. Campbell, S. Dauth, Downloaded from Survey on RGB, 3D, thermal, and multimodal regeneration with robotic implants. Sci. Robot. 3, A. K. Capulli, F. S. Pasqualini, S. Ahn, A. Cho, H. Yuan, approaches for facial expression recognition: History, eaaq0018 (2018). B. M. Maoz, R. Vijaykumar, J.-W. Choi, K. Deisseroth, trends, and affect-related applications. IEEE Trans. 94. B. J. Nelson, I. K. Kaliakatsos, J. J. Abbott, Microrobots for G. V. Lauder, L. Mahadevan, K. K. Parker, Phototactic Pattern Anal. Mach. Intell. 38, 1548–1568 (2016). minimally invasive medicine. Annu. Rev. Biomed. Eng. guidance of a tissue-engineered soft-robotic ray. 83. B. Scassellati, Theory of mind for a humanoid robot. 12, 55–85 (2010). Science 353, 158–162 (2016). Auton. Robot. 12, 13–24 (2002). 95. A. Servant, F. Qiu, M. Mazza, K. Kostarelos, B. J. Nelson, 106. H. Singh, T. Maksym, J. Wilkinson, G. Williams, 84. M. I. Jordan, T. M. Mitchell, Machine learning: Trends, Controlled in vivo swimming of a swarm of bacteria‐like Inexpensive, small AUVs for studying ice-covered polar http://robotics.sciencemag.org/ perspectives, and prospects. Science 349, 255–260 microrobotic flagella. Adv. Mater. 27, 2981–2988 environments. Sci. Robot. 2, eaan4809 (2017). (2015). (2015). 85. G.-Z. Yang, J. Cambias, K. Cleary, E. Daimler, J. Drake, 96. O. Felfoul, M. Mohammadi, S. Taherkhani, Acknowledgments: We thank all those who have responded P. E. Dupont, N. Hata, P. Kazanzides, S. Martel, R. V. Patel, D. de Lanauze, Y. Zhong Xu, D. Loghin, S. Essa, S. Jancik, to our online survey. In particular, we thank H. Choset, V. J. Santos, R. H. Taylor, Medical robotics—Regulatory, D. Houle, M. Lafleur, L. Gaboury, M. Tabrizian, N. Kaou, P. Dario, P. Fiorini, T. Fukuda, D. Leff, and S. Russell, who ethical, and legal considerations for increasing levels of M. Atkin, T. Vuong, G. Batist, N. Beauchemin, provided detailed comments, review, and original materials Sci. Robot. 2 autonomy. , eaam8638 (2017). D. Radzioch, S. Martel, Magneto-aerotactic bacteria for this article. 86. C. Bergeles, G.-Z. Yang, From passive tool holders to deliver drug-containing nanoliposomes to tumour microsurgeons: Safer, smaller, smarter surgical robots. hypoxic regions. Nat. Nanotechnol. 11, 941–947 10.1126/scirobotics.aar7650 IEEE Trans. Biomed. Eng. 61, 1565–1576 (2014). (2016). 87. V. Vitiello, S.-L. Lee, T. P. Cundy, G.-Z. Yang, Emerging 97. C. Cath, S. Wachter, B. Mittelstadt, M. Taddeo, L. Floridi, Citation: G.-Z. Yang, J. Bellingham, P. E. Dupont, P. Fischer, robotic platforms for minimally invasive surgery. IEEE Artificial Intelligence and the ‘Good Society’: The US, EU, L. Floridi, R. Full, N. Jacobstein, V. Kumar, M. McNutt, R. Merrifield, by guest on February 1, 2018 Rev. Biomed. Eng. 6, 111–126 (2013). and UK approach. Sci. Eng. Ethics 2017, 1–24 (2017). B. J. Nelson, B. Scassellati, M. Taddeo, R. Taylor, M. Veloso, 88. A. Shademan, R. S. Decker, J. D. Opfermann, S. Leonard, 98. L. Floridi, M. Taddeo, What is data ethics? Philos. Trans. A Z. L. Wang, R. Wood, The grand challenges of Science Robotics. A. Krieger, P. C. W. Kim, Supervised autonomous Math. Phys. Eng. Sci. 374, 20160360 (2016). Sci. Robot. 3, eaar7650 (2018).

Yang et al., Sci. Robot. 3, eaar7650 (2018) 31 January 2018 14 of 14 The grand challenges of Science Robotics Guang-Zhong Yang, Jim Bellingham, Pierre E. Dupont, Peer Fischer, Luciano Floridi, Robert Full, Neil Jacobstein, Vijay Kumar, Marcia McNutt, Robert Merrifield, Bradley J. Nelson, Brian Scassellati, Mariarosaria Taddeo, Russell Taylor, Manuela Veloso, Zhong Lin Wang and Robert Wood

Sci. Robotics 3, eaar7650. DOI: 10.1126/scirobotics.aar7650 Downloaded from

ARTICLE TOOLS http://robotics.sciencemag.org/content/3/14/eaar7650 http://robotics.sciencemag.org/

REFERENCES This article cites 89 articles, 14 of which you can access for free http://robotics.sciencemag.org/content/3/14/eaar7650#BIBL

PERMISSIONS http://www.sciencemag.org/help/reprints-and-permissions by guest on February 1, 2018

Use of this article is subject to the Terms of Service

Science Robotics (ISSN 2470-9476) is published by the American Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. 2017 © The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original U.S. Government Works. The title Science Robotics is a registered trademark of AAAS.