机器人英文论文及翻译

1、 Behavior RobotIntroductionAs a design strategy, the behavior-based approach has produced intelligent systems for use in a wide variety of areas, including military applications, mining, space exploration, agriculture, factory automation, service industries, waste management, health care, disaster i

2、ntervention and the home. To understand what behavior-based robotics is, it may be helpful to explain what it is not. The behavior-based approach does not necessarily seek to produce cognition or a human-like thinking process. While these aims are admirable, they can be misleading. Blaise Pascal onc

3、e pointed out the dangers A Nomad robot used by many researchers to study behavior within a laboratory setting. inherent when any system tries to model itself. It is natural for humans to model their own intelligence. The problem is that we are not aware of the myriad internal processes that actuall

4、y produce our intelligence, but rather experience the emergent phenomenon of thought. In the mid-eighties, Rodney Brooks (1986) recognized this fundamental problem and responded with one of the first well-formulated methodologies of the behavior-based approach. His underlying assertion was that cogn

5、ition is a chimera contrived by an observer who is necessarily biased by his/her own perspective on the environment. (Brooks 1991) As an entirely subjective fabrication of the observer, cognition cannot be measured or modeled scientifically. Even researchers who did not believe the phenomenon of cog

6、nition to be entirely illusory, admitted that AI had failed to produce it. Although many hope for a future when intelligent systems will be able to model human-like behavior accurately, they insist that this high-level behavior must be allowed to emerge from layers of control built from the bottom u

7、p. While some skeptics argue that a strict behavioral approach could never scale up to human modes of intelligence, others argued that the bottom-up behavioral approach is the very principle underlying all biological intelligence. (Brooks 1990) To many, this theoretical question simply was not the i

8、ssue. Instead of focusing on designing systems that could think intelligently, the emphasis had changed to creating agents that could Junior: An all-terrain robot recently used to deploy a gamma-locating device within a radioactive environment. act intelligently. From an engineering point of view, t

9、his change rejuvenated robotic design, producing physical robots that could accomplish real-world tasks without being told exactly how to do them. From a scientific point of view, researchers could now avoid high-level, armchair discussions about intelligence. Instead, intelligence could be assessed

10、 more objectively as a measurement of rational behavior on some task. Since successful completion of a task was now the goal, researchers no longer focused on designing elaborate processing systems and instead tried to make the coupling between perception and action as direct as possible. This aim r

11、emains the distinguishing characteristic of behavior-based robotics. The sub-sections which follow explain the roots of behavior based robotics, how it rose as a counter to the symbolic, deliberative approach of classical AI and how it has come to be a standard approach for developing autonomous rob

12、ots. A special thanks to Ronald Arkin whose book, Behavior Based Robotics, has greatly influenced this report. Understanding the Context of Classical AIClassical AI spent decades trying to model human-like intelligence, using knowledge-based systems that processed representation at a high, symbolic

13、level. Symbolic representation was considered of paramount importance because it allowed agents to operate on sophisticated human concepts and report on their action at a linguistic level. As Donald Michie stated, In AI-type learning, explainability is all. (Michie 1988) Since the goal of early AI w

14、as to produce human-like intelligence, researchers used human-like approaches. Marvin Minsky, in many ways a father of the field of AI, believed an intelligent machine should, like a human, first build a model of its environment and then explore solutions abstractly before enacting strategies in the

15、 real world. (McCarthy et al. 1955) This emphasis on symbolic representation and planning had a great effect on robotics and spurred control strategies where functionality was coded using languages and programming architectures that made conceptual sense to a human designer. Although many of the str

16、ategies developed were both elaborate and elegant, the problem was that the intelligence in these systems belonged to the designer. The robot itself had little or no autonomy and often failed to perform if the environment changed. While classical AI viewed intelligence as the ability of a program to

17、 process internal encodings, a behavior-based approach considers intelligence to be demonstrated through meaningful and purposeful action in an environment. (Arkin 1999) While many perceived the behavior-based movement to have forsaken the goal of human-like intelligence, others maintained that high

18、-level intelligence would indeed arise once a strong, low-level foundation had been laid. Agre and Chapman argued that, in fact, human beings are actually much more reactive than we imagine ourselves to be. (Agre and Chapman 1987) The planning and cognition that we are consciously aware of represent

19、s only the tip of a cerebral iceberg comprised mostly of unconscious, reactive motor skills and implicit behavior encodings. In a sense, the behavioral approach did not abandon modeling human intelligence as much as human consciousness. One of the sideeffects has been that many behavior-based approa

20、ches produce systems that are anything but explainable. High scientific aims aside, a main reason the behavior-based community is so intent on developing automated learning techniques is that a human designer often finds it excruciatingly tedious or impossibly difficult to orchestrate many behaviors

21、 operating in parallel. It is worse than frustrating to debug behavior that emerges from the interplay of many layers of asynchronous control. At times, a truly well-implemented, behavior-based approach will result in successful strategies the researchers themselves cannot explain or understand. Rea

22、ctive vs. DeliberativeThere is still considerable debate over the optimal role of internal representation. (Clark & Grush 1999) Many researchers believe that a robot cannot assign meaning to its actions or environment without representing them, even if indirectly. (Pylyshyn 1987) (Fodor 1987) Others

23、 believe that reliance on internal representation thwarts a robot抯 ability to act quickly across domains. An important figure for the field of behavior-based robotics, Rodney Brooks, declared planning to be just a way of avoiding figuring out what to do next. (Brooks 1987) Strategies which require t

24、hat action be mediated by some symbolic representation of the environment are often called deliberative. In contrast, reactive strategies do not exhibit a steadfast reliance on internal models, but displace some of the role of representation onto the environment itself. Instead of responding to enti

25、ties within a model, the robot can respond directly to perception of the real world. Thus, reactive systems are best characterized by a direct connection between sensors and effectors. Control is not mediated by a model but rather occurs as a low level pairing between stimulus and response. If a tas

26、k is highly structured and predictable it may make sense to use a deliberative approach. For example, if an intelligent agent is embedded in an entirely virtual environment, then it is often possible to encode every aspect of the environment with some semantic representation. In complex, real-world

27、domains where uncertainty cannot be effectively modeled, however, robots must have a means of reacting to an infinite number of possibilities. Some behavior-based strategies use no explicit model of the environment. In the late 1980抯 Schoppers believed that if a programmer knew enough about an envir

28、onment, s/he could make a set of stimulus-response pairs sufficient to cover every possibility. (Schoppers 1989) Clearly, such an approach is only possible in restricted domains such as a chess game or micro-world where there are a limited number of possible states. For more complicated domains it i

29、s necessary to find an appropriate balance between reactive and deliberative control. Systems that seek to completely avoid internal representation are ill-equipped for the many tasks that require memory or communication. On the other hand, systems that must transmute all perception and action throu

30、gh an internal model will be necessarily confounded in some new environment. The key is that the model should not drive development. Rather, control should be built from the bottom up and distributed across the system. For a reactive design methodology to work, it is necessary that behavior be decom

31、posed into atomistic components. Often, design will include a developmental phase during which these components can be honed and joined together. First, the designer builds a minimal system and then exercises it, using an ongoing loop to evaluate performance and add new competence. A Basis in Biolog

32、yBehavior-based robotics uses biology as the best model for understanding intelligence. Most roboticists do not model biological organisms directly, but rather look to nature for insight and direction. Increasingly, researchers have adopted the notion that high-level cognition is an impractical, deb

33、ilitating goal, and have begun to model the lower animal world. While there is a definite danger of trying to stretch metaphors too thin, the fact is that biological models offer our best hope for creating adaptive behavior. Biology serves not only as inspiration for underlying methodologies, but al

34、so for actual robot hardware and sensors. At the Centre National de la Recherche Scientifique in France, researchers discovered that a simple household fly navigates using a compound eye comprised of 3,000 facets which operate in parallel to monitor visual motion. In response, roboticists built an a

35、rtificial robot eye with 100 facets that can provide a 360-degree panoramic view. (Pranceshini, Pichon and Blanes 1992) Artificial bees can simulate the dance patterns and sounds of real bees sufficiently well to actually communicate with other bees. (Kirchner & Town 1994) Others have managed to bui

36、ld robot cockroaches (Quinn and Espenschied 1993) and even ants capable of leaving and detecting pheromone trails. (Russell, Thiel, & Mackay-Sim 1994). ARIEL: a behavior-based robot developed by iRobot to locate and detonate mines within the surf-zone. It is possible to view these successes as evide

37、nce supporting the behavior-based approach. In other words, if most animals do not rely on cognition to act, why should robotics? Roboticists preoccupation with high-level semantic thought merely reflects the anthropomorphic bias of human designers. To better understand the behavioral architecture o

38、f a low-level animal, scientists severed the connection between a frog抯 spine and brain. The goal was to remove all centralized control so that all action was produced reactively and without thought. Scientists stimulated particular points along the spinal cord and found much of the behavior of a fr

39、og was encoded directly into the spine. There are twenty locations along the spine, each of which can react with a different, essential motion. Stimulating one location will prompt the frog to wipe its head whereas another will cause it to jump. If the spine is stimulated in two points simultaneousl

40、y it is possible to combine behaviors and produce a more complex form of behavior. (Bizzi, Mussa-Ivaldi, & Giszter 1991) INAT: A robot developed at the Idaho National Engineering and Environmental Laboratory which uses learned responses to light and sound fluctuations to modulate swarming behavior.

41、This finding bears out a fundamental premise of the behavior-based approach: that sophisticated, high-level behavior can emerge from layered combinations of simple stimulus-response mappings. Instead of careful planning based on modeling, high-level behavior such as flocking or foraging can be built

42、 by blending low-level behaviors such as dispersion, aggregation, homing and wandering. Strategies can be built directly from behaviors, whereas plans must be based on an accurate model. DART: An aquatic robot developed by iRobot. Of course, it is not only Biology which supplies insight to the field

43、 of robotics. As multi-disciplinary approaches become more prevalent, inspiration should flow freely between robotics, neuroscience, psychology, cognitive science, and a host of other fields. Control ArchitectureNow that we have explained why reactive control is useful, it remains to be shown how th

44、is reactive control is actually accomplished. There must be some control architecture that puts these conceptual ideas to work. Maya Mataric defines the purpose of architecture as a principled way of organizing a control system, and further explains that, in addition to providing structure, it impos

45、es constraints on the way the control problem can be solved. (Mataric 1992) Early researchers focused on planning modules and, because many of the agents operated in a virtual world, de-emphasized the part of the architecture that controlled the motors and sensors. This section explains the architec

46、tures that were adopted in opposition to this mindset. The new architectures constrained development by forcing a distributed approach where behaviors function in parallel rather than in a step-wise, linear fashion. Later, it explains how hybrid approaches attempt to reintegrate some of the old aims

47、 of deliberative, cognitive techniques back into the behavior-based approach. Subsumption Architectures The subsumption architecture originally developed by Brooks in 1986 provided a method for structuring reactive systems from the bottom up using layered sets of rules. (Brooks 1986) Bottom-layer be

48、haviors such as avoid-collision should be the most basic and should have the highest priority. Top-layer behaviors such as go to goal encapsulate high-level-intention and may be built from lower behaviors or may function only when lower behaviors such as avoid collision are satisfied. To reduce comp

49、lexity, there should be minimal interaction between behaviors. The idea is that each should function simultaneously but asynchronously with no dependence on the others. This independence should reduce interference between behaviors and prevent overcomplexity. Successes using subsumption architecture

50、s include six-legged walking robots, vacuuming agents and robots that collect cans. The layered approach promotes the fault tolerance and robustness necessary for such agents. For instance, although a robot designer cannot accurately predict component failure, a well-designed subsumption architectur

51、e will allow behaviors to sequence and re-sequence according to unforeseen problems. Subsumption architectures do not require an explicit plan of action. Clearly, one of the biggest challenges in designing a subsumption architecture is giving the system the ability to automatically select among beha

52、viors (arbitration). While the ability to assign priorities to behaviors already affords some organization, it is necessary to have a program that can smoothly transition between states. For instance, a robot that has finished vacuuming a room must end some behaviors and begin some others before it

53、can exit and dump its trash. The situated automata strategy developed by Rosenschein and Kaelbling in 1990 produced a reactive system where finite state automata are used to identify and react to discrete events. Other possibilities include a winner-take-all approach where spreading activation conve

54、rges to a specific behavior or a voting architecture such as DAMN, where each behavior has some voice in deciding a single output. Hybrid Control SystemsSubsumption architectures are inherently reactive. They intend to enable real-time control, robust performance, modular development, learning throu

55、gh incremental growth, bottom-up design and a tight coupling between action and incoming sensory data. For some tasks, however, there is an inescapable need for willed rather than automatic control. Although subsumption architectures seek to avoid representation, it can be very useful to link symbol

56、ic meaning to action. Representation allows the designer to explicitly evoke intentionality from the robot. Deliberative systems can readily integrate world knowledge into a system, whereas purely reactive systems cannot. Ideally, there should be some way to combine the benefits of a reactive strate

57、gy with a mechanism that can be aware of its environment and able to communicate about it. Hybrid approaches have attempted to make planning subordinate to reactivity and yet still use it to guide reactivity at a high level. Arkin believes that reactive and deliberative control can be complementary:

58、 The false dichotomy that exists between hierarchical control and reactive systems should be dropped. (Arkin 1989) In fact, planning can be viewed as a configuration for selecting behaviors. Arkin designed an autonomous robot architecture (AURA) that used representation-based, hierarchical component

59、s including a mission planner, spatial planner and plan sequencer to advise a reactive component. (Arkin 1986) NASA rovers such as Robby use a hybrid architecture that could acknowledge failure and adapt the reactive controller accordingly. (Gat 1992) Lyons used a planner to produce continuous modification of a reactive system according to some high-level goal. (Lyons 1992) In the same year, Connell developed a system with a servo layer, subsumptio