SCIENCE robotics 2024年4月24日文章:为什么动物跑得比机器人快
添加图片注释，不超过 140 字（可选）Abstract 摘要
Animals are much better at running than robots. The difference in performance arises in the important dimensions of agility, range, and robustness. To understand the underlying causes for this performance gap, we compare natural and artificial technologies in the five subsystems critical for running: power, frame, actuation, sensing, and control. With few exceptions, engineering technologies meet or exceed the performance of their biological counterparts. We conclude that biology’s advantage over engineering arises from better integration of subsystems, and we identify four fundamental obstacles that roboticists must overcome. Toward this goal, we highlight promising research directions that have outsized potential to help future running robots achieve animal-level performance.动物比机器人更擅长跑步。性能上的差异体现在敏捷性、范围和健壮性等重要方面。为了理解造成这种性能差距的根本原因，我们比较了对运行至关重要的五个子系统中的自然技术和人工技术:动力、框架、驱动、传感和控制。除了少数例外，工程技术达到或超过其生物对应物的性能。我们的结论是，生物学相对于工程学的优势来自于子系统更好的整合，我们确定了机器人专家必须克服的四个基本障碍。为了实现这一目标，我们强调了有前途的研究方向，这些方向具有巨大的潜力，可以帮助未来的跑步机器人实现动物水平的表现。
Animals outperform robots at locomotion. The performance gap is evident across scales, and it is particularly galling given that animal designs respect constraints that need not limit robots; for instance, animals must grow from a single cell, repair their own bodies, and contain all the machinery needed to reproduce. We seek to understand the underlying causes for this performance gap by systematically comparing animals with robots.动物在运动方面胜过机器人。性能差距在各个尺度上都很明显，考虑到动物设计尊重不需要限制机器人的约束，这尤其令人恼火;例如，动物必须从一个单细胞生长，修复自己的身体，并包含繁殖所需的所有机械。我们试图通过系统地比较动物和机器人来了解这种表现差距的潜在原因。
Although the preceding observations apply to multiple locomotion modalities, including flight and swimming, for tractability, we focus on legged locomotion, where decades of research have produced a rich robot ecosystem with biocomparable designs. For succinctness and to emphasize high-performance behavior, we will use the catchall phrase “runner” to refer to animals and robots that use intermittent contact between limbs and terrain to move and “running” to refer to the corresponding behavior, regardless of whether it would be more common or accurate to describe a behavior as walking or jumping. Toward these ends, we seek to answer the question, “Why can animals outrun robots?”尽管前面的观察结果适用于多种运动模式，包括飞行和游泳，但我们关注的是腿运动，几十年的研究已经产生了丰富的机器人生态系统，具有生物可比性设计。为了简洁和强调高性能的行为，我们将使用“runner”这个笼统的短语来指代利用四肢和地形之间的间歇性接触来移动的动物和机器人，而“running”则指代相应的行为，而不管用“walking”或“jumping”来描述一种行为是否更常见或更准确。为了达到这些目的，我们试图回答这个问题:“为什么动物能跑得比机器人快?”
Our goal is motivated, in part, by bioinspiration and biomimetic approaches to design (1–3), that is, the potential to advance robotics by translating natural to artificial technology, as well as robot-inspired approaches to biology (4) and physics (5), wherein robots are used to advance basic science. Quantifying the performance of “proof-of-concept” designs embodied by extant animals sets aspirational benchmarks for the robotics community, highlights performance limiters, and potentially reveals design principles. We expect that this study will help catalyze advancements in bioinspired and biohybrid robotics and extremes of performance achievable by autonomous robots (6).我们的目标在一定程度上受到生物灵感和仿生设计方法的推动(-)，即通过将自然技术转化为人工技术来推进机器人技术的潜力，以及机器人灵感的生物学()和物理学()方法，其中机器人用于推进基础科学。量化现存动物体现的“概念验证”设计的性能，为机器人社区设定了理想的基准，突出了性能限制，并潜在地揭示了设计原则。我们期望这项研究将有助于促进生物启发和生物混合机器人技术的进步，以及自主机器人可实现的极限性能()。
Engineered and biological runners are built differently. Robots are assembled from discrete components at the macroscale, whereas animals are formed from heterogeneous structures grown at the nanoscale. Additionally, the two technologies use different physical phenomena and materials for power, sensing, actuation, and control. However, both animals and robots are built to run (among other tasks). Given that this shared objective is achieved using vastly different design paradigms, it is not obvious how to compare animal and robot runners. Thus, we consider multiple levels of analysis (7, 8), first by quantifying the performance gap between the systems as a whole, as in Fig. 1, and subsequently by comparing performance across the five subsystems critical for the task of running illustrated in Fig. 2. Last, we conclude by synthesizing our findings to propose fruitful future directions for running robot research.工程赛跑者和生物赛跑者的构造不同。机器人是由宏观尺度上离散的部件组装而成，而动物是由纳米尺度上生长的异质结构形成的。此外，这两种技术在动力、传感、驱动和控制方面使用了不同的物理现象和材料。然而，动物和机器人都是为奔跑而生的(在其他任务中)。考虑到这一共同目标是使用截然不同的设计范例实现的，如何比较动物和机器人跑步者并不明显。因此，我们考虑多个层次的分析(，)，首先通过将系统之间的性能差距作为一个整体进行量化，如图1所示，然后通过比较图2所示的运行任务的五个关键子系统的性能。最后，综合本文的研究结果，提出了未来跑步机器人研究的方向。
Fig. 1. System-level performance of animal and robot runners.图1所示。动物和机器人跑步者的系统级性能。(A) Representative performance of robots (blue) and animals (orange) in the three-dimensional space defined by range, agility, and robustness axes. (B) Projection of (A) onto agi
(A)机器人(蓝色)和动物(橙色)在由距离、敏捷性和鲁棒性轴定义的三维空间中的代表性表现。(B) (A)在敏捷范围平面上的投影。(C) (A)在敏捷-鲁棒平面上的投影。动物的跑步表现现在在所有尺度上都是帕累托式的。来源:a. mastin
图2所示。五个子系统对运行至关重要。
(A) Block diagram showing interconnections among the power, frame, actuation, sensing, and control subsystems as they interact with the environment (orange, yellow, red, blue, light red, and black, respectively). Solid arrows indicate transduction of force or energy, and dashed arrows indicate transmission of information. (B) Illustration of the five subsystems overlaid on the fastest running animal (cheetah): fat and metabolism; bone skeleton; muscles; visual, vestibular, and proprioceptive sensors; nervous system. (C) Illustration of the five subsystems overlaid on the fastest autonomous running robot (WildCat): gas engine or electric battery; metal or carbon fiber struts; hydraulic or (piezo)electric motors; vision, IMU, and joint sensors; computer network.
CREDIT: A. MASTIN/SCIENCE ROBOTICS
(A)显示电源、框架、驱动、传感和控制子系统与环境交互时的互连框图(分别为橙色、黄色、红色、蓝色、浅红色和黑色)。实线箭头表示力或能量的传递，虚线箭头表示信息的传递。(B)覆盖在跑得最快的动物(猎豹)上的五个子系统示意图:脂肪和新陈代谢;骨骨架;肌肉;视觉、前庭和本体感觉传感器;神经系统。(C)最快自主运行机器人(WildCat)上覆盖的五个子系统示意图:燃气发动机或电池;金属或碳纤维支撑;液压或(压电)电动机;视觉、IMU和关节传感器;计算机网络。
In the following subsections, we compare performance measurements from the literature on animal physiology and robot design. The metrics that we choose are largely scale invariant, at least above a minimum size where engineered systems struggle, and are measured in diverse taxa, including vertebrates, invertebrates, and robots. We exclude technologies not integrated into existing autonomous runners, for instance: Spider silk is very strong, but it is not used as a structural material in animal locomotion; nuclear reactors can power submarines but have not been integrated into running robots.在下面的小节中，我们比较了动物生理学和机器人设计方面的文献中的性能测量。我们选择的度量标准在很大程度上是尺度不变的，至少在工程系统难以达到的最小尺寸之上，并且在不同的分类群中进行测量，包括脊椎动物、无脊椎动物和机器人。我们排除了没有集成到现有自动跑步器中的技术，例如:蜘蛛丝非常坚固，但它没有被用作动物运动的结构材料;核反应堆可以为潜艇提供动力，但尚未集成到运行机器人中。
A true meta-analysis remains out of reach because we found no principled weighting by which the performance of such diverse organisms and machines could be distilled into an average value. Instead, we selected data from representative systems to informally assess whether and how biological components exceed the performance of their engineered counterparts; these data are summarized in Fig. 3. Because a comprehensive metric table with citations would consume this manuscript, we present the details in a supplementary document. We encourage the reader not to skip the supplement but instead read it for a deeper look at the component metrics, data, and rationale underpinning our assessments.真正的荟萃分析仍然遥不可及，因为我们没有发现有原则的权重，可以将这些不同生物和机器的表现提炼成一个平均值。相反，我们从代表性系统中选择数据，非正式地评估生物成分是否以及如何超过其工程对应物的性能;这些数据汇总在图3中。因为一个包含引文的综合度量表会消耗这份手稿，所以我们在补充文档中提供了详细信息。我们鼓励读者不要跳过补充部分，而是更深入地了解支撑我们评估的组成指标、数据和基本原理。
Fig. 3. Subsystem-level performance of animal and robot runners above 1 kg.图3所示。1公斤以上的动物和机器人跑者的子系统级表现。Performance in each subsystem is compared using multiple performance metrics and one or more engineering technologies. The highest-performing quantities
SYSTEM PERFORMANCE 系统性能
Although the claim that animals outperform robots at running may sound uncontroversial in 2024, it is, nevertheless, worthwhile to consider how to quantify the performance gap. We think that a runner should have range to operate independently over the distances required, agility to reach and traverse surfaces in its environment, and robustness to maintain range and agility despite changes to the runner and its environment. Although running performance could be measured along other axes, these three nonredundant metrics are commonly studied and of paramount importance for animal fitness and robot autonomy (9).尽管在2024年，动物在跑步方面胜过机器人的说法听起来毫无争议，但考虑如何量化这种表现差距是值得的。我们认为，一个跑步者应该有足够的距离来独立完成所需的距离，有足够的灵活性来到达和穿越环境中的表面，有足够的稳健性来保持距离和灵活性，尽管跑步者和环境发生了变化。尽管可以沿着其他轴来测量跑步性能，但这三个非冗余指标通常被研究，并且对动物适应性和机器人自主性至关重要()。
Range can be directly quantified as the distance traveled during autonomous running in a specific environment. This distance is determined by the onboard energy supply as well as the efficiency of energy conversion. The latter factor is conventionally measured by the cost of transport, defined as the amount of energy required to move a unit weight of a runner over a unit distance (10). The farthest walk by a legged robot on a single battery charge was Ranger’s 65-km trek over the course of 31 hours (11). Ranger’s cost of transport is impressively half that of human walking, but there are important caveats. First, the robot’s batteries have about 50-fold less useful energy per unit mass as compared with animal fat; the average human has energy reserves to continue walking long after the robot’s batteries are depleted. When allowed to refuel along the way, humans can exhibit extraordinary endurance: Exceptional athletes can run hundreds of kilometers over multiple days in a single outing. They can also do so over rough terrain, whereas Ranger exploits the smoothness of the track that it was designed to walk on—a small rock could cause it to stumble and fall. Outside of controlled environments, robot range is a distant second to that of animals.里程可以直接量化为在特定环境中自动行驶时所走的距离。这个距离是由机载能量供应以及能量转换的效率决定的。后一个因素通常是通过运输成本来衡量的，运输成本定义为将一个单位重量的跑步者移动单位距离()所需的能量。有腿机器人在一次充电的情况下行走的最远距离是“游侠”(Ranger)在31小时内行走了65公里。游侠的运输成本是人类步行的一半，但有一些重要的注意事项。首先，与动物脂肪相比，机器人电池每单位质量的有用能量少了大约50倍;在机器人的电池耗尽后，一般人都有能量储备，可以继续行走很长时间。当被允许在途中补充能量时，人类可以表现出非凡的耐力:杰出的运动员可以在一次郊游中连续数天跑数百公里。它们也可以在崎岖的地形上行走，而“游侠”则利用了它设计时所使用的轨道的平稳性——一块小岩石可能会使它绊倒并摔倒。在受控环境之外，机器人的活动范围远远落后于动物。
The agility of legged robots has been quantified using running speed, jumping height, turning rate, and more. Legged robot development has long been guided by a need for speed, resulting in bipeds, quadrupeds, and hexapods with speeds approaching those of similar-sized animals on regular terrain (12–19). However, animals are still faster at all scales, and the performance gap widens when considering irregular or deformable terrain (20, 21). Some legged robots have leaped at the task of jumping, either specializing entirely (22) or by adapting an existing runner (23). New heights have been reached by legged robots using bioinspired elastic energy storage, but even these are still surpassed by animals of similar mass. Rapid robot turning has been occasionally synthesized (24, 25), but animals can redirect momentum “on a dime” (26–31). Last, although feedback control can enable robots to recover from substantial perturbations (32), the ability of animals in this regard is unmatched (33). Overall, animals out-run and out-maneuver running robots.有腿机器人的敏捷性已经用奔跑速度、跳跃高度、转弯速度等来量化。长期以来，有腿机器人的发展一直以对速度的需求为指导，导致了两足动物、四足动物和六足动物在正常地形上的速度接近。然而，动物在所有尺度上仍然更快，当考虑不规则或变形地形时，性能差距会扩大(，)。一些有腿的机器人已经完成了跳跃的任务，要么是完全专业化的()，要么是通过改造现有的跑步者()。利用生物弹性能量储存的有腿机器人已经达到了新的高度，但即使这样，同样质量的动物仍然超过了它们。机器人的快速转弯偶尔被合成(，)，但动物可以“瞬间”改变动量(-)。最后，虽然反馈控制可以使机器人从严重的扰动中恢复()，但动物在这方面的能力是无法比拟的()。总的来说，动物跑得比机器人快，动作也比机器人灵活。
Robustness is deceptively easy to conceptualize yet devilishly difficult to quantify. As a starting point, we consider how agility and range are maintained in the presence of changes to the runner or its environment. Horses can increase body mass 20-fold as they grow from foal to full size, and rhinoceros beetles can carry 30 times their body weight without fatiguing. Animals can survive bone fracture (34) or limb loss, with many lizards and insects voluntarily shedding appendages to distract predators (35), and the phenomenon is so common in the latter group as to motivate a “spare leg hypothesis” (36). In contrast, robot range and agility decrease precipitously when large payloads are added or limbs are damaged. Robots designed to walk or run on flat ground can be made to plod over rough terrain under inclement conditions (37), but animals are unimpeded by terrain variations upward of their height (21, 38–41) and readily run over, under, and through obstacles like mud, snow, vegetation, rubble, and crevices (42, 43). On granular media, robot running speed can depend sensitively on design, control, and environmental parameters (44), with animals handily outpacing robots in their native ecologies (20). Overall, animals excel at maintaining performance despite changes that would be catastrophic for existing robots.稳健性看似很容易概念化，但却极难量化。作为起点，我们考虑如何在跑步者或其环境发生变化的情况下保持敏捷性和范围。马在从马驹到成年的过程中，体重可以增加20倍，而犀牛甲虫可以负重30倍于体重的物体而不会感到疲劳。动物可以在骨折()或断肢()中存活下来，许多蜥蜴和昆虫自愿脱落附属物以分散捕食者的注意力()，这种现象在后一组中非常普遍，以至于激发了“多余的腿假说”()。相反，当增加大载荷或肢体损坏时，机器人的范围和敏捷性急剧下降。设计用于在平地上行走或奔跑的机器人可以在恶劣的条件下在崎岖的地形上行走()，但动物不受地形变化的影响(，-)，并且可以轻松地在泥泞，雪，植被，碎石和裂缝等障碍物上，下面和通过(，)。在颗粒介质中，机器人的运行速度可以敏感地依赖于设计、控制和环境参数()，动物在其原生生态中很容易超过机器人()。总的来说，动物在保持性能方面表现出色，尽管这些变化对现有机器人来说可能是灾难性的。
We conclude that animals outperform robots at running along the three key axes of range, agility, and robustness, as illustrated in Fig. 1 and corroborated in other recent work (45). In what follows, we seek to understand the cause of this performance gap. Given that animals and robots are generally designed and built using different technologies, it is possible that differences in the parts give rise to differences in the whole. To test this hypothesis, we coarsely divided runners into the five subsystems illustrated in Fig. 2: a power system to store and deliver energy; a frame for support and leverage; actuators to modulate mechanical energy; sensors to perceive self and environment; and a control system to transmit and transform sensor and actuator signals. Of course, our “subsystems” are abstractions, and runners cannot always be cleanly divided, particularly in the case of animals. We will take care in what follows to note when separating subsystems is messy work.我们得出结论，动物在跑动的范围、敏捷性和稳健性这三个关键轴上优于机器人，如图1所示，并在最近的其他研究中得到证实()。接下来，我们将试图了解这种绩效差距的原因。考虑到动物和机器人通常是用不同的技术设计和制造的，部分的差异可能会导致整体的差异。为了验证这一假设，我们将运行器大致分为图2所示的五个子系统:存储和传递能量的电力系统;支撑和杠杆的框架;调节机械能的执行器;感知自我和环境的传感器;控制系统用于传感器和执行器信号的传输和转换。当然，我们的“子系统”是抽象的，赛跑者不能总是被清晰地划分，特别是在动物的情况下。当分离子系统是一项棘手的工作时，我们将注意下面要注意的事项。
POWER SUBSYSTEM PERFORMANCE电源子系统性能
The ideal power supply for running stores a large amount of useful energy and delivers it efficiently to the other subsystems with minimal added mass. The three main types of power plants used in autonomous runners are gas engines, electric batteries, and metabolism. All three convert stored chemical energy to power running: Engines convert gas to movement, batteries convert chemical bonds to electricity, and metabolism converts fat to adenosine triphosphate (ATP).理想的运行电源存储大量的有用能量，并以最小的附加质量将其有效地传递给其他子系统。自动跑者使用的三种主要动力装置是燃气发动机、电池和新陈代谢装置。这三者都将储存的化学能转化为动力:发动机将气体转化为运动，电池将化学键转化为电能，新陈代谢将脂肪转化为三磷酸腺苷(ATP)。
Because a runner’s endurance is ultimately limited by its stored energy, we compare mass-specific stored energy, defined as the energy delivered by the power plant normalized by fuel mass. Biology outperforms engineering by this metric, with values more than double those of combustion engines and 50-fold more than batteries. There are two main reasons for biology’s edge: Oxidative metabolism within mitochondria converts fat to ATP with a remarkable efficiency of about 70% (46) compared with 25% in engines (47); and, whereas adipose tissue is almost 90% fuel (48), gas tanks can be 20% of the mass of the fuel they carry (49).因为跑步者的耐力最终受到其储存能量的限制，我们比较了质量比储存能量，定义为由燃料质量标准化的发电厂提供的能量。按照这个标准，生物学的价值超过工程学，是内燃机的两倍多，是电池的50倍多。生物学的优势有两个主要原因:线粒体内的氧化代谢将脂肪转化为ATP，其效率约为70%()，而发动机的效率为25% ();而且，脂肪组织几乎是90%的燃料()，而油箱可以是其所携带燃料质量的20%()。
Because locomotion is among the most power-intensive behaviors that runners perform, we compare mass-specific delivered power, defined as the sustainable power delivered normalized by total power plant mass. Metabolism meets or exceeds engine performance by this metric (47, 50), but batteries outperform both using the natively high power output of lithium-ion cells and relatively light electronics and packaging. Although animals may transiently achieve higher peak power outputs by depleting the supply of ATP in muscles, the energy in stored ATP is quite limited and, if used on its own, could only sustain performance for a few seconds (51, 52).由于运动是跑步者进行的最耗能的行为之一，我们比较了质量特定的传递功率，定义为由发电厂总质量标准化的可持续传递功率。新陈代谢达到或超过了发动机的性能(，)，但电池的性能优于锂离子电池本身的高功率输出和相对较轻的电子和包装。虽然动物可以通过消耗肌肉中ATP的供应来短暂地获得更高的峰值功率输出，但储存在ATP中的能量是相当有限的，如果单独使用，只能维持几秒钟的表现(，)。
Because fuel can potentially be harvested from the environment to extend running distances, we compare mass-specific refueling power, defined as the energy rate of refueling divided by the mass of the refueling frame. By this metric, gas tanks can be refueled an order of magnitude faster than a battery can charge or digestion can process biological matter. To put this in perspective, a human would only need to refuel at the gas rate for a fraction of a second to gain the energy it needs for each day. The actual human refueling rate limits 100-day running to a range of about 40 km per day (53).由于燃料有可能从环境中获得以延长行驶距离，我们比较了质量比加油功率，定义为加油的能量率除以加油框架的质量。根据这个标准，油箱的加油速度比电池充电速度快一个数量级，消化系统处理生物物质的速度也快一个数量级。从这个角度来看，一个人只需要以汽油的速度补充几分之一秒的能量就可以获得每天所需的能量。实际的人类加油速度限制了100天的跑步，每天大约40公里()。
In summary, the performance of engineered power plants can exceed that of their biological counterparts in the rates at which they refuel and deliver energy, although biology now has the edge in energy storage. The development of portable power plants capable of delivering both high specific energy and high specific power is considered one of the grand challenges for mobile robots (6). Fortunately, there is no known fundamental barrier to creating engineered power plants that have a superior combination of storage and energy-delivering capabilities (54).总之，工程发电厂的性能在补充燃料和输送能量的速度上可以超过生物发电厂，尽管生物发电厂现在在能量储存方面具有优势。开发能够同时提供高比能和高比功率的便携式发电厂被认为是移动机器人面临的重大挑战之一()。幸运的是，目前还没有什么已知的根本障碍，来创造具有储存和输送能量能力的卓越组合的工程发电厂()。
RAME SUBSYSTEM PERFORMANCE框架子系统性能
The ideal frame for running combines material and geometry to support and propel the body overground while being light and failure resistant. Running robot frames are generally built from rigid connections between steel, aluminum, or carbon fiber struts using linear or rotary joints. Animal frames have two primary forms: Vertebrates have an endoskeleton made from bone connected by soft tissue, and insects have an exoskeleton made from hard cuticle connected by soft flexures (55). A runner’s frame is loaded by multiaxial forces that vary over time (56), making it susceptible to multiple modes of failure, including buckling and yielding. Failure modes are affected by the frame’s geometric and material properties, and bracing against one failure mode may weaken the frame against another (57). A simple yet instructive analysis is to consider a macroshape shared between robot and animal frames (55), a cylindrical tube, and evaluate material resistance to failure modes dominated by stiffness and strength.理想的跑步框架结合了材料和几何形状，以支持和推动身体在地面上，同时轻便和抗故障。跑步机器人的框架通常由钢、铝或碳纤维支柱之间的刚性连接构成，使用线性或旋转关节。动物的骨架主要有两种形式:脊椎动物的内骨骼由骨头和软组织连接而成，昆虫的外骨骼由坚硬的角质层和柔软的弯曲连接而成()。转轮车架受到随时间变化的多轴力()的载荷，使其容易受到多种失效模式的影响，包括屈曲和屈服。破坏模式受框架的几何和材料特性的影响，支撑对抗一种破坏模式可能会削弱框架对抗另一种破坏模式()。一个简单但有指导意义的分析是考虑机器人和动物框架之间共享的宏观形状()，圆柱形管，并评估材料对由刚度和强度主导的破坏模式的抗力。
Because the body’s weight must be supported throughout running without buckling, we compare density-specific stiffness, defined as a material’s modulus of elasticity normalized by its density. Carbon fiber outperforms the other materials by this metric by a factor of 3 to 5, with cuticle, bone, aluminum, and steel being roughly comparable. Because limbs must generate large forces to propel the body overground without breaking, we compare density-specific strength, defined as a material’s stress before fracture normalized by its density. Carbon fiber again outperforms the other materials by this metric by an order of magnitude; the substantial density of steel makes it the lowest performer in the group. In practical terms, this means that a carbon fiber limb could support a heavier body and enable more agile maneuvers that would otherwise break a bone, fracture an exoskeleton, or snap a strut made of aluminum or steel with similar mass.因为在整个运行过程中，车身的重量必须在不弯曲的情况下得到支撑，所以我们比较了密度比刚度，密度比刚度定义为材料的弹性模量由其密度归一化。在这方面，碳纤维的性能比其他材料高出3到5倍，而角质层、骨骼、铝和钢的性能大致相当。因为四肢必须产生巨大的力量来推动身体在地面上不断裂，所以我们比较了密度比强度，定义为材料在断裂前的应力按其密度归一化。碳纤维在这个指标上又比其他材料好了一个数量级;钢的高密度使其成为性能最差的钢。在实际应用中，这意味着碳纤维肢体可以支撑更重的身体，并实现更灵活的动作，否则就会折断骨头，折断外骨骼，或者折断类似质量的铝或钢制成的支柱。
Because frames and joints often store and return energy, we compare mass-specific energy, defined as a material’s strength squared normalized by its stiffness and density. In this metric, carbon fiber outperforms the other materials above by a factor of 3 to 10. However, there are other materials, both engineered and biological, that have exceptional specific energy that can be used for the sole purpose of storing and returning energy. For example, resilin is used in insect jumpers (58) and tendon in vertebrates (59). Both have higher specific energy than carbon fiber, but not higher than what is achievable by engineered rubber and Kevlar (60).因为框架和关节经常储存和返回能量，我们比较质量比能，定义为材料的强度平方归一化的刚度和密度。在这个指标中，碳纤维的性能比上述其他材料高出3到10倍。然而，还有其他材料，无论是工程材料还是生物材料，它们具有特殊的特异能量，可以用于存储和返回能量的唯一目的。例如，弹性蛋白用于昆虫的跳跃()和脊椎动物的肌腱()。两者都比碳纤维具有更高的比能，但并不高于工程橡胶和凯夫拉(Kevlar)所能达到的比能。
In summary, engineered frames built from carbon fiber can be much stiffer and stronger than biological skeletons built from bone, cuticle, aluminum, or steel; metals may outperform biology with respect to stiffness but underperform in strength. Animal frames now exhibit a greater diversity of microshapes (for instance, trabecular bones have remarkable crack propagation resistance) that offer advantages beyond our metrics (61). However, a growing catalog of materials and fabrication techniques available to robots may provide similar advantages (6, 62).总之，由碳纤维制成的工程框架比由骨头、角质层、铝或钢制成的生物骨架要坚固得多;金属在硬度方面可能优于生物，但在强度方面表现不佳。动物骨架现在表现出更大的微形状多样性(例如，小梁骨具有显著的抗裂纹扩展能力)，这提供了超出我们指标的优势()。然而，越来越多的材料和制造技术可以为机器人提供类似的优势(，)。
ACTUATION SUBSYSTEM PERFORMANCE驱动子系统性能
The ideal actuators for running enable rapid changes in runner momentum with minimal added mass. Animal runners exclusively actuate their limbs with muscle, and most autonomous running robots use electromagnetic motors at vertebrate scales or piezoelectrics at insect scales. The physical principles governing motors, piezos, and muscle are different: Motors produce force from the flow of current in a magnetic field, piezos use crystal properties to convert electric fields to mechanical pressure, and muscles produce force through chemical reactions that generate length changes in nanoscale proteins. We exclude other actuators that have been deployed in running robots, including hydraulics and artificial muscles (63), because motors and piezos are sufficient to justify our quantitative conclusions below.理想的运行执行器能够以最小的附加质量快速改变流道动量。动物跑步者只用肌肉驱动四肢，大多数自主跑步的机器人在脊椎动物层面上使用电磁马达，在昆虫层面上使用压电马达。控制马达、压电片和肌肉的物理原理是不同的:马达通过磁场中的电流产生力，压电片利用晶体特性将电场转化为机械压力，而肌肉通过化学反应产生力，从而在纳米级蛋白质中产生长度变化。我们排除了在运行机器人中部署的其他执行器，包括液压和人造肌肉()，因为马达和压电足以证明我们下面的定量结论是正确的。
Because running requires high forces to support the body and move the limbs, we compare mass-specific peak torque, defined as the maximum torque normalized by the mass of the actuator and its transmission. Muscles can outperform direct-drive motors and bimorph piezos (64) by a factor of 2 to 5 in this metric. Although transmissions can theoretically multiply torques by arbitrarily high gear ratios, the mass added and efficiency lost yield diminishing returns in this metric. Nevertheless, the performance gap between motors and muscles can be eliminated at the vertebrate scale by pairing motors with higher-ratio transmissions, like harmonic drives (65) or ball screws (66), and series compliance to provide backdrivability (67).因为跑步需要很大的力量来支撑身体和移动四肢，我们比较了质量相关的峰值扭矩，定义为执行器及其传动的质量标准化的最大扭矩。在这个指标中，肌肉可以比直接驱动电机和双变形压电()高出2到5倍。虽然变速器理论上可以将扭矩乘以任意高的传动比，但在这一度量中，质量的增加和效率的损失导致收益递减。然而，在脊椎动物的尺度上，电机和肌肉之间的性能差距可以通过将电机与更高传动比的传动装置配对来消除，比如谐波传动()或滚珠丝杠传动()，并通过串联来提供反向驱动能力()。
Because running agility is limited by the rate at which actuator output can be converted to a change in momentum of the runner, we compare mass-specific power, defined as the mean mechanical power over a gait cycle normalized by the mass of the actuator and its transmission. By this metric, peak performance of EM motors exceeds that of muscle by one or more orders of magnitude, and piezos are comparable to muscle when used for running. Sustained performance in motors, piezos, and muscles alike is constrained by thermal management and energy supply. Although animals use spring-assisted power amplification to overcome actuator limitations (68), robots can also use these mechanisms (22).由于跑步敏捷性受限于驱动器输出转换为跑步者动量变化的速率，我们比较了质量比功率，定义为由驱动器质量及其传动归一化的步态周期的平均机械功率。通过这个指标，EM电机的峰值性能超过肌肉的一个或多个数量级，并且压电在用于跑步时可与肌肉相媲美。电机、压电片和肌肉的持续性能都受到热管理和能量供应的限制。虽然动物使用弹簧辅助功率放大来克服执行器的限制()，但机器人也可以使用这些机制()。
In summary, the performance of motors with high-ratio transmissions and series compliance can meet or exceed that of muscles in torque and power density, whereas piezos only match muscle in the latter and are at a disadvantage in the former. Motors and piezos have an advantage over muscles in their efficiency of energy transduction, which can exceed 90%, whereas muscle fiber efficiency in most animals is closer to 30% . Hydraulic actuators may exceed the torque and power density of motors and muscles, but their efficiencies are often much lower than either, and they require a complex and heavy fluid system in parallel with the electrical systems used for sensing and control. Natural muscles’ variable shapes and inherent scalability provide packaging advantages not available in motors, easily adding degrees of freedom where needed, distributing actuation mass elegantly across the body, and providing failure tolerance through redundancy. The diverse linear actuator technologies known as “artificial muscles” may offer similar advantages but now have an equally diverse set of limitations compared to motors.综上所述，具有高传动比和串联柔度的电机在扭矩和功率密度上可以达到或超过肌肉的性能，而压电电机仅在扭矩和功率密度上与肌肉相匹配，在串联柔度上处于劣势。马达和压电在能量传导效率方面比肌肉有优势，可以超过90%，而大多数动物的肌纤维效率接近30%[在最极端的情况下测量不到63%;())。液压致动器的扭矩和功率密度可能超过马达和肌肉，但它们的效率往往比两者都低得多，而且它们需要一个复杂而沉重的流体系统，与用于传感和控制的电气系统并行。天然肌肉的可变形状和固有的可扩展性提供了电机无法提供的封装优势，在需要时轻松增加自由度，在整个身体上优雅地分配驱动质量，并通过冗余提供故障容忍度。被称为“人造肌肉”的各种线性执行器技术可能提供类似的优势，但与电机相比，现在具有同样多样化的限制。
SENSING SUBSYSTEM PERFORMANCE传感子系统性能
The ideal sensor suite for running delivers the actionable information (70) needed to move quickly overground. There are two fundamental sensing modalities relevant to running: electromagnetic and mechanical. Eyes, cameras, and LIDAR (light detection and ranging) are examples of the former; vestibular systems, inertial measurement units, and force transducers are examples of the latter. The mechanistic details differ in biology and engineering: Animals sense light via chemical excitation from photon absorption and strain via ion channels that open in response to the physical deformation of membranes or molecules; robots sense light via electrical excitation from photon absorption and mechanical deformations via strain, magnetic fields, and optics. However, both vision and mechanosensation technologies generally transduce sensory cues into analog electrical signals that are subsequently encoded into digital signals.理想的传感器套件可以提供在地面上快速移动所需的可操作信息。有两种与跑步相关的基本传感方式:电磁和机械。眼睛、相机和激光雷达(光探测和测距)是前者的例子;前庭系统、惯性测量单元和力传感器是后者的例子。机理细节在生物学和工程学上有所不同:动物通过光子吸收产生的化学激发来感知光，并通过响应膜或分子的物理变形而打开的离子通道进行应变;机器人通过光子吸收产生的电激发和通过应变、磁场和光学产生的机械变形来感知光。然而，视觉和机械感觉技术通常都是将感觉信号转换成模拟电信号，然后编码成数字信号。
Because the information provided by a sensor is fundamentally limited by its ability to perceive change in the world, we compare threshold sensitivity, defined as the smallest unit of input that results in a resolved response from the sensor. Both biological and engineered sensors can nearly achieve theoretical limits: single photons and microstrains. For example, biological photoreceptors can resolve individual “quantum bumps” of electrical activity from absorption of single photons (71), similar to single-photon avalanche diodes in semiconductors (72). The ubiquitous invertebrate campaniform sensilla can detect strains as small as proteins (73), whereas hair cells in the mammalian vestibular and auditory system go further still (74). However, engineered strain gauges can be many orders of magnitude more sensitive (75).由于传感器提供的信息基本上受到其感知世界变化的能力的限制，因此我们比较了阈值灵敏度，阈值灵敏度定义为传感器产生解析响应的最小输入单位。生物和工程传感器几乎都能达到理论极限:单光子和微应变。例如，生物光感受器可以通过吸收单光子()来解决电活动的单个“量子颠簸”，类似于半导体中的单光子雪崩二极管()。无所不在的无脊椎动物的钟形感受器可以检测到小到蛋白质的品系()，而哺乳动物前庭和听觉系统中的毛细胞则更进一步()。然而，工程应变片的灵敏度可以提高许多数量级()。
Robots generally use a handful of sensors, whereas animals have large numbers distributed throughout their bodies. Because redundant distributed sensors can yield richer data more robustly, we compare the number of sensors in each modality. The number of rod and cone cells in a human eye (74) is comparable to the number of pixels in the latest smartphones. Invertebrate compound eyes have far fewer individual receptors, with cockroaches having comparable numbers to LIDAR arrays (76). However, animals have many orders of magnitude more strain sensors than robots. Humans, for instance, have roughly 200,000 tactile receptors in addition to 50,000 stretch receptors (77, 78). Insects can have thousands of individual campaniform sensilla to detect exoskeleton strains, thousands of mechanosensitive neurons in chordotonal organs that detect internal strains, and hundreds to thousands of other less well-characterized mechanosensory hairs and sensilla (79).机器人通常使用少量传感器，而动物则有大量分布在全身的传感器。由于冗余分布式传感器可以更鲁棒地产生更丰富的数据，因此我们比较了每种模式下传感器的数量。人眼中的视杆细胞和视锥细胞的数量相当于最新智能手机的像素数量。无脊椎动物复眼的单个感受器要少得多，蟑螂的感受器数量与激光雷达阵列相当()。然而，动物的应变传感器比机器人多出许多数量级。例如，人类大约有20万个触觉感受器和5万个拉伸感受器(，)。昆虫可以有数千个单独的钟形感受器来检测外骨骼菌株，在chordotonal器官中有数千个机械敏感神经元来检测内部菌株，以及数百到数千个其他不太清楚的机械感觉毛发和感受器()。
In summary, biological and engineered photoreceptors are comparable in their overall counts and ability to detect visual stimuli. Although engineered mechanoreceptors can detect much smaller stimuli than biological ones (74, 75), biology’s ability to integrate staggering numbers of mechanosensors distributed throughout bodies, including the electrical system needed to innervate the sensors, is remarkable. There are robustness advantages to the redundant mechanosensing streams in animals, given that failure of any particular sensor need not halt a runner in its tracks. Further, the ability to sense throughout the body may also confer advantages for agility, because actionable information may arise in any nook or cranny.总之，生物和工程光感受器在它们的总数量和检测视觉刺激的能力上是相当的。尽管工程机械感受器可以探测到比生物感受器小得多的刺激，但生物将分布在人体各处的数量惊人的机械感受器整合在一起的能力，包括控制这些感受器所需的电系统，是非常了不起的。动物体内的冗余机械传感流具有鲁棒性优势，因为任何特定传感器的故障都不需要停止奔跑。此外，全身感知的能力也可能赋予敏捷性优势，因为可操作的信息可能出现在任何角落或缝隙中。
Another potentially interesting comparison is the cost associated with sensing. In animals, neural activity in sensory regions can be substantial (for instance, 8% of resting metabolic rate for the blowfly retina), placing evolutionary pressure on the size and processing of nervous systems (80, 81). However, during movement, metabolic rate increases up to 50-fold (82). Together, these observations imply that the contribution of sensing to the overall energy budget during running is low and therefore is not predictive of overall performance differences, even if they might still be important in the overall fitness of the animal.另一个可能有趣的比较是与传感相关的成本。在动物中，感觉区域的神经活动可能很大(例如，苍蝇视网膜的静息代谢率为8%)，这对神经系统的大小和处理施加了进化压力(，)。然而，在运动过程中，代谢率增加了50倍()。总之，这些观察表明，在跑步过程中，感知对整体能量预算的贡献很低，因此不能预测整体性能差异，即使它们在动物的整体健康中可能仍然很重要。
CONTROL SUBSYSTEM PERFORMANCE控制子系统性能
The ideal controller for running transmits and transforms sensor and actuator signals to produce versatile behavior. Although animals can walk in the absence of large parts of their nervous systems (83) and robots can walk without computers (84), electrical control systems are used to run. The physical components and mechanisms differ in biological and engineered control systems: Neurons transmit action potentials through axons and synapses using the diffusion of charged molecules; electrical circuits and networks transmit binary or analog signals on wires using electromagnetic waves. Because implementing a control policy requires communication and computation, we consider both in what follows. For the former, we compare axons to network cables; for the latter, we compare natural and artificial spiking neural networks. Larger runners have more time to react to sensor signals before they hit the ground, so we normalize time by the natural period of a runner’s limb.理想的控制器运行传输和转换传感器和执行器信号，以产生多功能的行为。虽然动物可以在没有大部分神经系统的情况下行走，机器人也可以在没有计算机的情况下行走，但电气控制系统是用来运行的。生物和工程控制系统的物理成分和机制不同:神经元通过轴突和突触传递动作电位，利用带电分子的扩散;电路和网络利用电磁波在导线上传输二进制或模拟信号。由于实现控制策略需要通信和计算，因此我们将在以下内容中考虑这两者。对于前者，我们将轴突比作网线;对于后者，我们比较了自然和人工尖峰神经网络。体型较大的跑步者在落地前有更多的时间对传感器信号做出反应，所以我们用跑步者肢体的自然周期来规范时间。
Bandwidth and latency fundamentally limit controller performance (85), so we compare both. Many axons can be bundled into a single nerve to increase bandwidth without affecting latency, so we normalize bandwidth by the cross-sectional area of the communication channel. The bandwidth of a standard Ethernet cable at 10 megabit/s is at least 10,000 times greater than the fastest single axon, but a bundle of 1 million human axons has comparable area- and period-specific bandwidth (86), whereas gigabit Ethernet and other computer network protocols are orders of magnitude faster still. In addition, period-specific latency is at least 1000 times longer in nerves than an Ethernet cable, and it is impractical for biology to close this gap (87). At the smallest scales, buses connecting integrated circuits can have orders of magnitude higher bandwidth and lower latency than Ethernet and, thus, even greater advantages over axons.带宽和延迟从根本上限制了控制器的性能()，因此我们对两者进行比较。许多轴突可以捆绑在单个神经中以增加带宽而不影响延迟，因此我们通过通信通道的横截面积对带宽进行归一化。10兆比特/秒的标准以太网电缆的带宽至少是最快的单个轴突的1万倍，但是100万个人类轴突的束具有类似的特定区域和特定周期的带宽，而千兆以太网和其他计算机网络协议的速度要快得多。此外，神经的周期特异性延迟至少比以太网电缆长1000倍，从生物学角度来看，要缩小这一差距是不切实际的()。在最小的尺度上，连接集成电路的总线可以比以太网具有更高的带宽和更低的延迟，因此，比轴突具有更大的优势。
Effective controllers quickly compute complex policies. The time required for computation in spiking neural networks is proportional to the period-specific latency of a neuron, the time constant of which is on the order of milliseconds for natural neurons (88) and shorter than microseconds for artificial neurons (89). However, the number of neurons and synapses differs vastly in natural and artificial networks, with biology outperforming engineering in this metric by orders of magnitude at scales ranging from insects to people. However, it is worth remembering that animals rely on their nervous systems to implement a rich repertoire of behaviors, including attracting a mate, finding food, and avoiding predators. It is unclear how much brain is needed for locomotion, given that parasitic wasps with fewer than 400 neurons can fly, feed, and find hosts (90).有效的控制器可以快速计算复杂的策略。在尖峰神经网络中，计算所需的时间与神经元的特定周期延迟成正比，自然神经元()的时间常数在毫秒量级，而人工神经元()的时间常数小于微秒。然而，在自然网络和人工网络中，神经元和突触的数量差别很大，从昆虫到人类，生物学在这一指标上的表现要比工程学好几个数量级。然而，值得记住的是，动物依靠它们的神经系统来实现丰富的行为，包括吸引配偶、寻找食物和躲避捕食者。目前尚不清楚运动需要多少大脑，因为只有不到400个神经元的寄生蜂可以飞行、觅食和寻找宿主()。
In summary, computer networks vastly exceed the performance of nervous systems in latency and bandwidth of communication and computation, but artificial neural networks are at a substantial disadvantage relative to the size and connectivity of biological networks. Animals cannot practically decrease sensorimotor delay by the orders of magnitude that would be required to compete with robots’ communication channels; this fundamental limit surely affects control strategies, for instance, by favoring the use of internal models (87). Although neuromorphic circuits will continue to increase in complexity, it remains to be seen whether bigger brains are better (91) for running and how to make most effective use of the limited brainpower available to robots in the meantime (92).总之，计算机网络在通信和计算的延迟和带宽方面大大超过了神经系统的性能，但相对于生物网络的大小和连通性，人工神经网络处于实质性的劣势。实际上，动物无法将感觉运动延迟减少到与机器人的通信通道竞争所需的数量级;这个基本限制肯定会影响控制策略，例如，通过支持使用内部模型()。尽管神经形态回路的复杂性将继续增加，但是否更大的大脑更适合跑步，以及如何在此期间最有效地利用机器人有限的脑力，仍有待观察。
DISCUSSION 讨论
Returning to the hypothesis posed at the outset, we found some limited evidence, summarized in Fig. 3, that performance differences at the level of subsystems favor biology, partly explaining why animals outrun robots. Fat stores a lot of energy per unit mass, giving animals an advantage in range, particularly compared with robots powered by batteries. Muscle has higher torque density than piezos and motors paired with conventional transmissions, likely conferring some advantage in agility. Although biological sensors are no more sensitive than their engineered counterparts, a large number of them can be distributed throughout the body, lending robustness through redundancy and benefitting agility by providing rich sensor streams from each body part. Last, brains can theoretically implement much more complex transformations than current integrated circuits because of their vastly greater quantities of neurons and synapses, potentially leading to advantages in range, agility, and robustness. Biological subsystems fare better with respect to robots at insect scales than at human scales, indicating substantial headroom for component performance improvements in roach-sized runners.回到一开始提出的假设，我们发现了一些有限的证据，总结在图3中，子系统水平上的性能差异有利于生物学，部分解释了为什么动物比机器人跑得快。每单位质量的脂肪储存了大量的能量，使动物在射程上具有优势，特别是与电池驱动的机器人相比。肌肉的扭矩密度比压电和马达配上传统变速器要高，这可能会在敏捷性方面带来一些优势。虽然生物传感器并不比工程传感器更敏感，但大量的生物传感器可以分布在全身，通过冗余来增强鲁棒性，并通过提供来自每个身体部位的丰富传感器流来增强敏捷性。最后，从理论上讲，大脑可以实现比当前集成电路更复杂的转换，因为它们的神经元和突触数量要大得多，这可能会在范围、敏捷性和稳健性方面带来优势。生物子系统在昆虫尺度上比在人类尺度上表现得更好，这表明在蟑螂大小的跑步者中，组件性能有很大的改进空间。
However, a simple thought experiment demonstrates that these differences in runner parts do not explain most of the gap in running performance. Suppose we could build cyborg runners using the highest-performing components and subsystems from biological and engineering technologies: a fat-burning, carbon-fibered, muscle-bound monstrosity with distributed sensors and low-latency engineered communication channels, all controlled by mind-bogglingly complex spiking neural networks. Would roboticists be able to create cyborgs whose running performance competes with those of animals? This experiment could be carried out in the world of computational simulations, where runner designs are not constrained by the innumerable practical obstacles that make our imagined cyborg physically unrealizable. Even in the most favorable of these worlds, where frames never break and nearly unlimited computational resources control ideal torque sources using perfect state information, we suspect that the performance of simulated runners would not approach the agility or robustness of animals in the real world.然而，一个简单的思维实验表明，这些跑步者部位的差异并不能解释跑步表现上的大部分差距。假设我们可以使用生物和工程技术中性能最高的组件和子系统来建造半机械人跑步者:一个燃烧脂肪、碳纤维、肌肉发达的怪物，拥有分布式传感器和低延迟的工程通信通道，所有这些都由令人难以置信的复杂脉冲神经网络控制。机器人专家能否创造出跑步性能堪比动物的半机械人?这个实验可以在计算模拟的世界中进行，在这个世界中，跑步者的设计不受无数实际障碍的限制，这些障碍使我们想象中的半机械人在物理上无法实现。即使在这些最有利的世界中，帧永远不会中断，几乎无限的计算资源使用完美的状态信息控制理想的扭矩源，我们怀疑模拟跑步者的表现不会接近现实世界中动物的敏捷性或鲁棒性。
If not the performance of subsystems, what is the explanation for why animals can outrun robots? By elimination, the problem must lie with our lack of understanding of how to construct and control a high-performance “whole” using existing high-performance “parts.” This is a forgivable shortcoming because at least four fundamental obstacles must be overcome to tackle this integration challenge. First, we lack quantitative metrics for evaluating the many dimensions of running performance, yet these are necessary for improving robot designs using systematic engineering processes. We qualitatively discussed agility and robustness at the outset, but there are only a handful of narrowly defined ways to measure these properties. Even range, which we conflated with distance, is only well-defined once the runner’s behavior and environment are specified. The second and third obstacles are trade-offs and emergence. Stringent trade-offs potentially arise when subsystems combine because performance of one component might constrain that of another. The opposite is also possible during integration, because emergence is where the behavior of the whole is different than, and irreducible to, the behavior of the parts. The composition of subsystems, especially when feedback is involved, can transform the dynamics for better or worse. These two obstacles are two sides of the same coin in the sense that, at their core, they are unknown but potentially transformative interaction dynamics and that performing the integration is the only way to expose these dynamics. However, there are a huge number of ways that the parts can be combined, each producing different possibilities for trade-offs and emergent behavior. Unfortunately, the fourth obstacle, the curse of dimensionality (93, 94), admonishes us that these high-dimensional integration spaces cannot be explored by brute force alone. Consequently, it is challenging to find good mechanical designs in the high-dimensional space of candidate designs and good control policies in the high-dimensional space of candidate policies.如果不是子系统的性能，那么动物跑得比机器人快的解释是什么呢?通过消除，问题一定在于我们缺乏对如何使用现有的高性能“部件”构建和控制高性能“整体”的理解。这是一个可以原谅的缺点，因为要解决这个集成挑战，至少必须克服四个基本障碍。首先，我们缺乏定量指标来评估运行性能的许多维度，然而这些对于使用系统工程流程改进机器人设计是必要的。我们从一开始就定性地讨论了敏捷性和健壮性，但是只有少数几个狭义定义的方法来度量这些属性。甚至距离(我们将其与距离混为一谈)也只有在指定了跑步者的行为和环境后才能定义清楚。第二和第三个障碍是权衡和涌现。当子系统组合在一起时，可能会出现严格的权衡，因为一个组件的性能可能会约束另一个组件的性能。在整合过程中也可能出现相反的情况，因为涌现是整体的行为不同于部分的行为，并且不可简化为部分的行为。子系统的组成，特别是当涉及到反馈时，可以更好或更坏地改变动态。这两个障碍是同一枚硬币的两面，从某种意义上说，它们的核心是未知的，但可能具有变革性的交互动态，而执行集成是暴露这些动态的唯一方法。然而，有很多方法可以将这些部分组合在一起，每种方法都会产生不同的权衡和紧急行为的可能性。不幸的是，第四个障碍，维度的诅咒(，)告诫我们，这些高维的集成空间不能仅仅通过蛮力来探索。因此，在候选设计的高维空间中找到好的机械设计，在候选策略的高维空间中找到好的控制策略是一项挑战。
How can the daunting challenges of integration be overcome? Given that tackling the entire system-level problem is daunting, decomposing into subproblems is helpful. The conventional subsystems that we evaluate above are one such decomposition. However, performance in these subsystems has been driven by industry’s need to efficiently manufacture at scale rather than the roboticist’s desire to build the ideal subsystem for running. Instead, we advocate for decomposing into “functional subunits”: groupings of parts that reveal the trade-offs and emergent behavior arising from their integration. As an example, consider the series elastic joint actuator (95, 96): Composed of elements of frame, sensing, actuation, and low-level control, its design is subject to the integration challenges and trade-offs that we argue are central to the performance deficit of running robots. At the same time, it features emergent behavior greater than the sum of its parts, because it is torque- and power-dense while maintaining backdrivability and robust force control, simplifying high-level control. The complexity of functional subunits should be more tractable than that of a whole robot, enabling tight integration and performance optimization for their subtasks. They should have reduced and predictable ways of interacting mechanically and electrically with other functional subunits to simplify integration into the broader system. The reduced subset of possibilities ought to make the overall design space more feasible to navigate while still allowing a rich set of runners to explore. As a final note, functional subunit decomposition is compatible with proven tools for building and analyzing runners. For instance, hierarchical models of varying degrees of complexity (97, 98) have revealed how reduced-order emergent behavior is embodied in more complex machines (99, 100)—functional subunits could facilitate this embodiment. Additionally, information-based metrics of control architectures like centralization (101) and control effort (102, 103) provide potential design criteria at an integrative level applicable to functional subunits and whole robots alike. Although we believe that this approach of decomposing the problem of runner design into functional subunit design will be fruitful, we also understand that it will require creativity, inspiration, and discovery.如何才能克服一体化的艰巨挑战?考虑到处理整个系统级问题是令人生畏的，将其分解成子问题是有帮助的。我们上面评估的常规子系统就是这样一个分解。然而，这些子系统的性能是由工业对大规模高效制造的需求驱动的，而不是机器人专家对构建理想运行子系统的渴望。相反，我们主张将其分解为“功能子单元”:揭示其集成所产生的权衡和紧急行为的部分分组。例如，考虑系列弹性关节驱动器(，):由框架，传感，驱动和低级控制元件组成，其设计受到集成挑战和权衡的影响，我们认为这是运行机器人性能缺陷的核心。与此同时，它的紧急行为大于其各部分的总和，因为它是扭矩和功率密集的，同时保持了反驾驶性和强大的力控制，简化了高级控制。功能子单元的复杂性应该比整个机器人的复杂性更易于处理，从而实现子任务的紧密集成和性能优化。它们应该具有减少和可预测的与其他功能子单元进行机械和电气交互的方式，以简化集成到更广泛的系统中。减少的可能性子集应该使整体设计空间更加可行，同时仍然允许丰富的跑步者进行探索。最后需要说明的是，功能子单元分解与构建和分析运行程序的经过验证的工具是兼容的。例如，不同复杂程度的层次模型(，)揭示了如何在更复杂的机器中体现低阶紧急行为(，)-功能子单元可以促进这种体现。此外，控制体系结构的基于信息的度量，如集中化()和控制努力()，提供了在集成级别适用于功能子单元和整个机器人的潜在设计标准。虽然我们相信这种将跑步道设计问题分解为功能亚单元设计的方法将是富有成效的，但我们也明白，这需要创造力、灵感和发现。
Lest roboticists feel sheepish about their machines’ performance, we note that biology has a substantial head start over engineering to explore design and policy spaces. At the lineage level, there have been 1000 to 10,000,000 times as many generations of animals as robots. Considering population size, there have been 1000 times more humans than robots (of all kinds) and perhaps 1 quintillion times as many individual insects. In terms of individual experience, animals are less sedentary and have longer lifespans than robots, with ambulatory adult humans taking roughly 10,000 steps per day over decades. In light of these observations, it strikes us that the rate of advancements has been markedly faster in robots than in animals.为了避免机器人专家对他们的机器表现感到害羞，我们注意到，在探索设计和政策空间方面，生物学比工程学有很大的领先优势。在谱系水平上，动物的世代数是机器人的1000到1000万倍。考虑到人口规模，人类比机器人(各种机器人)多1000倍，个体昆虫的数量可能是机器人的1万亿倍。就个体体验而言，动物比机器人更少久坐不动，寿命更长，走动的成年人几十年来每天大约走1万步。根据这些观察结果，我们惊讶地发现，机器人的进步速度明显快于动物。
There are several key factors contributing to the disparity between the pace of technological innovation in biology and engineering. Designing and prototyping in engineering is a rapid and systematic process compared with the undirected search of evolution. Animals must survive to pass on their genes, limiting experimentation from generation to generation. Additionally, animals in one phylogenetic branch generally cannot benefit from innovations in any other: An adaptation that improves running in a cockroach provides no benefit to a cheetah. In contrast, advancements demonstrated in one robot are readily transferred to others. Furthermore, robots have access to sources of parallelism unavailable to animals: Experience can be accumulated on multiple physical and simulated robots simultaneously, and these data can be shared directly. Further, these advantages are only limited by the resources invested, for example, by the number of researchers in robotics labs, robots on the ground, or servers in the cloud.有几个关键因素导致了生物学和工程学技术创新速度的差异。与进化的无向搜索相比，工程中的设计和原型是一个快速而系统的过程。动物必须生存下来才能传递它们的基因，这就限制了实验代代相传。此外，一个系统发育分支中的动物通常不能从任何其他系统发育分支的创新中受益:提高蟑螂跑步能力的适应能力对猎豹没有任何好处。相反，在一个机器人上展示的进步很容易转移到其他机器人上。此外，机器人可以获得动物无法获得的并行性来源:可以同时在多个物理和模拟机器人上积累经验，并且这些数据可以直接共享。此外，这些优势只受投入资源的限制，例如，机器人实验室、地面机器人或云服务器中的研究人员数量。
We are optimistic that legged robots will someday outrun animals. To hasten this outcome, we conclude by highlighting emerging approaches that we regard as potentially transformative. The multidirectional exchange of principles and approaches among engineering, biology, and physics (1, 2, 4–6, 92) has yielded a wonderful constellation of insights and creative designs that have pushed the boundaries of knowledge and possibility. Going forward, systematic comparative studies (rather than single-species inspiration) could reveal generalizable principles for exceptional performance by providing evolutionary context for the factors shaping organisms (104). Distributing energy, sensing, actuation, and control throughout robot frames, as animals do, may advance autonomy (62, 105). Bridging the “sim-to-real” gap with better computational models of robot interaction with the environment (20, 44) could markedly accelerate exploration of design and policy spaces by reducing the number of physical prototypes that need to be built. Bodies can be made easier to control by offloading computation into morphology (84, 106); this approach remains underdeveloped, but continued advances in material robotics may prove transformative (62). Systematically exploring trade-offs with respect to multiple performance metrics promotes reuse of parts in disparate behaviors (6).我们乐观地认为，有腿的机器人有一天会超过动物。为了加速这一结果，我们总结了我们认为具有潜在变革性的新兴方法。工程学、生物学和物理学(，，-，)之间原理和方法的多向交流产生了一系列奇妙的见解和创造性的设计，这些设计推动了知识和可能性的界限。展望未来，系统的比较研究(而不是单一物种的启发)可以通过为形成生物体的因素提供进化背景，揭示出卓越表现的一般原则()。在整个机器人框架中分配能量、传感、驱动和控制，就像动物一样，可以提高自主性(，)。通过更好的机器人与环境交互的计算模型(，)来弥合“模拟到真实”的差距，可以通过减少需要建造的物理原型的数量，显著加快对设计和政策空间的探索。通过将计算卸载到形态学(，)，可以使物体更容易控制;这种方法仍然不发达，但材料机器人技术的持续进步可能会被证明是革命性的。系统地探索与多个性能度量相关的权衡，促进了不同行为中部件的重用()。
The lesson that we take from biology is that, although further improvements to components and subsystems are beneficial, the greatest opportunity to improve running robots is to make better use of existing parts. We advocate for integrative exploration of design and policy spaces.我们从生物学中得到的教训是，尽管进一步改进组件和子系统是有益的，但改进运行机器人的最大机会是更好地利用现有部件。我们提倡设计与政策空间的整合探索。
页:
[1]