"The robotics revolution is in an early stage, poised to explode at the dawn of the 21st century."
(Rodney A. Brooks)
"In the future, advanced robots may blur the difference between syntax and semantics, such that their responses will be indistinguishable from the responses of a human."(Michio Kaku)
"The best model of the world is the world itself" (Rodney A. Brooks).
Robotics
Robotics is the science and technology of robots, of the design, manufacture and applications of robots. It is a field that involves several disciplines: mechanics, electronics, computer science, artificial intelligence and control engineering.
The robot one of the cultural icons of the twentieth century are at the center of the debate on consciousness, artificial intelligence, life, cognitive science, linguistics and philosophy. They also allow us to test theories of knowledge or cognitive models.
The classical conception of a robot −that of classical artificial intelligence− is that of a machine with 3 components:
Sensors (for perception of the environment).
A control organ (the "mind" of the robot).
A motor device (in charge of performing certain physical actions on the environment).
This traditional conception was challenged by Rodney Brooks, who proposed a new artificial intelligence (AI) that he called "Nouvelle AI". Brooks published since 1986 a series of articles −including "Elephants don't play chess" [1990], "Intelligence without representation" [1991] and "Intelligence without reason" [1991]− which had a great impact and brought about a paradigm shift and a renewal of robotics. They moved from the computer model, with the control organ as the governing center of behavior (the model of artificial intelligence), to a biologically inspired model (Brooks was inspired by insects) based on a set of simple and coordinated individual behaviors, where the control organ is obviated and only considers perception and motor tasks. According to Brooks, intelligence has evolved from a single cell to today's humans, so with AI we must proceed in the same way: we must go from simple to complex and based only on the physical world. For Brooks, the most fascinating thing about biological evolution is that it raises questions about the limit or boundary between animate and inert matter.
Types of robots
There are many types of robots. The most important ones are:
Manipulator robots. They are multifunctional mechanical systems, which have a simple control system.
Control robots. They perform actions based on commands from the control module.
Learning robots. They repeat the movements performed by a human operator.
Metamorphic (or polymorphic) robots. They are flexible, self-managing, self-organizing and self-modifying at the structural and functional level. They are related to the subject of consciousness.
Cognitive (or intelligent) robots. Traditional robotics has been concerned with the design of robots to perform specific tasks. Cognitive robotics is inspired by cognitive science and studies robots with intelligent behavior, which learns, reasons and decides how to interact with the environment (what tasks to execute and at what time). There is a main agent (which plays the role of the robot's "mind") that is in charge of integrating and coordinating the different subsystems of the robot.
Cognition is the process of acquiring knowledge through the experience of interacting with the environment. Cognitive abilities are: planning, perception, attention, anticipation, general knowledge (beliefs) and specific knowledge.
Autonomous intelligent robots. The construction of an autonomous intelligent robot, with humanoid form and human capabilities, a recurring theme in science fiction, is considered the "Holy Grail" of artificial intelligence. If achieved, it will be the third industrial revolution. They will be mobile cognitive robots for the human environment: homes, streets, schools, hospitals, etc. They will have flexible behavior, they will reason, evolve and act like people. They will decide their behavior at all times. They will be able to internally represent the external world and able to move through the environment without stumbling. An autonomous intelligent robot project was Cog [see Addendum].
Hard robots. Robots with centralized control, pre-programmed movements, precise motors and a structured environment (no uncertainty).
Soft robots. Robots without centralized control that work in uncertain environments and with flexible behavior.
Sofbots. They are software robots, which "live" in a computer and evoke the idea of automated work. The science that studies them is often called "softbotics". Sensors and motors are implemented in software. An example of a softbot is Anna, IKEA's softbot, which is plays the role of expert and advisor.
Knowbots. They are softbots specialized in collecting and storing information in distributed databases, through data networks, to use it in a certain task, and to be able to share it with other processes or other softbots. If softbot means "automated work", knowbot means "automated knowledge".
A knowbot can be a resident program or travel the network, it can replicate itself to leave working copies where it has located information. There are specialized knowbots: on users' computers, on servers, traffic control, message transport, etc. An example of a softbot is the Web crawler developed by Google, which analyzes link structure and other content characteristics. Another example of a knowbot is the one that searches for people in lists of e-mail users on computers connected to the Internet.
The Paradigms of Robotics
The main paradigms of robotics are: hierarchical, reactive and hybrid.
The hierarchical paradigm
The hierarchical paradigm, also called deliberative or PPA (Perception - Planning - Action), is based on a chain of 3 sequential operations, in which each operation depends on the previous one, since it takes as input the outputs of the previous operation. Each operation is performed by a different module. This paradigm is inspired by the human model and artificial intelligence: first, it perceives; second, it "thinks" or decides what to do; and third, it acts. Of the three phases, the most important is the second, planning, which is the cognitive center of the robot.
The hierarchical paradigm is the classical paradigm, the oldest, and had its greatest boom in the 1960s and 1970s.
The reactive paradigm
It is the paradigm introduced by Brooks, when he considered that the traditional AI −based on the hierarchical model and whose initial objective was the emulation of the human mind− had failed because:
Intelligence is too complex, so the focus shifted to specific subproblems such as: language processing, machine vision, knowledge representation, etc.
AI had made dogma out of the model of abstract symbol manipulation, an imperfect and restrictive model that has produced poor results.
The hierarchical model placed too much emphasis on the cognitive, rational, planning aspect, which forced the creation of an internal intelligence model that was often difficult to realize.
Therefore, a new, different, radical, simpler and more practical paradigm was needed. This paradigm emerged in the 1980s, and its influence extended into the early 1990s.
Brooks realized that intelligence is not limited to abstract reasoning, that there is a simpler, more direct type of intelligence that required little analysis. The new AI proposed by Brooks is based on emphasizing purely physical interaction with the environment as the foundation of intelligent robotic systems. This paradigm is also called PA (Perception - Action) because it obviates planning and relies only on direct correspondences between perception and action. Intelligence comes from the binomial "perceive-act". First we have perception and then we have action. Instinct has no internal cognitive model. There are only automatic perception-action relationships.
This paradigm is also called "behavior-based AI", i.e., an external-type paradigm, which does not consider the internal, the "mental" of the robot. "Cognition is only in the eye of the beholder." The traditional AI position is that cognition comes first.
This paradigm is summarized in Brooks' famous article "Elephants don't play chess", by which his author meant that elephants do not need to perform symbolic or abstract reasoning to subsist in their environment. What is important is that the animal perceives the environment and acts on it in a reactive way.
Brooks justifies his model on the idea that many things we do are just interactions with the environment, rather than operating with a preconceived plan. "Much of what we do is completely beneath our awareness, and then we rationalize and explain what we have done."
Characteristics of the reactive paradigm:
It is a simpler paradigm, requiring fewer resources. It does not require a process of transforming the perceived into a symbolic structure. Perceptions are passed directly (or after minimal processing), in the form of subsymbolic inputs, to the action module that produces the external behavior.
It is based on starting with the simplest intelligence (such as that of insects) and then making more complex robots, in a bottom-up process in which the degree of intelligence is progressively increased. It is based on coordinating simple (low-level) mechanisms to produce complex behavior.
There is a subsumption architecture −"subsumed" means "included in something broader"− to model the behavior of autonomous robots by means of a set of modules. Each single individual module generates behavior.
A subsumption architecture is built by decomposing the system into subsystems (or layers) of independent behaviors (or minimal interaction). Each layer connects perception with action.
Each layer extracts only aspects of the world that are relevant and has its own purpose. The hardware of each layer is separate from the others.
Each layer constantly monitors the environment and responds with corresponding behavior.
Each layer of the architecture is a network of finite state machines, operating in parallel and asynchronously and sending messages to each other. The arrival of messages or the expiration of timers triggers the transition between states.
Each layer communicates with lower layers via "suppress" and "inhibit" messages to temporarily suspend their ongoing operation.
Brooks and his team built several robots to implement these ideas, including Allen, a robot developed in the late 1980s, which was the first robot based on the subsumption architecture. It had 3 layers: 1) avoid static and dynamic obstacles; 2) perform random motion every 10 seconds; 3) scan space to detect distant free space and approach it.
This model is used for moderately simple tasks that involve little reasoning. It is a practical intelligence that allows the robot to handle itself effectively in its environment. It has been the basis of successful applications in various fields: industry, agriculture, mining, domestic, entertainment, etc. The robot Mars Explorer was designed largely following the reactive model.
The hybrid paradigm
It combines the hierarchical and reactive paradigms. It is the most widely used paradigm and has been used since the 1990s until today. The hierarchical paradigm is applied in general or high-level planning. The reactive paradigm is used when direct and immediate action needs to be taken on the basis of sensor information.
Comparison Traditional AI vs. New AI
Comparison between both models, according to Brooks (1: traditional AI, 2: new AI)
Foundation.
It adopts an unsubstantiated hypothesis, which is that of equivalence between intelligence and computation.
It is based on a physical grounding hypothesis and, therefore, unquestionable. It is a fundamentalist, physical or situated AI.
Design.
It is based on the decomposition (by human intervention) of intelligence into functional modules of information processing. None of these modules alone generates behavior. It is necessary to combine all the modules to achieve the overall behavior. It is a top-down construction, from high-level abstractions to low-level or detail abstractions through a process of decomposition of the system into modules.
It is an incremental bottom-up construction. It starts with simple modules, which become progressively more sophisticated. At each step there is a complete system, with a certain degree of intelligence. There are no high-level abstractions, but high-level concretions based on low-level concretions. Complexity is not designed top-down, but arises from the interaction of simple components.
Control.
Requires centralized control.
No centralized control. Therefore, there are no bottlenecks and less likelihood of total collapse.
Competition.
To improve competence, individual functional modules must be improved.
Improving system competency is achieved by adding new simple individual modules. A small change can produce a large impact on the overall behavior.
Operational.
Intelligence works only with symbols and must be "fed" symbols by the perceptual system and must transform the symbols into signals for the motor system.
It does not need to use symbols nor does it need the classical inputs and outputs. It handles direct physical signals. It only needs a set of sensors (input signals) and actuators (output signals associated with actions).
Intelligence.
Assumes that intelligence is something objective and knowable.
Intelligence emerges from simple individual modules of behavior whose coexistence and cooperation with each other give rise to complex behavior and which is externally observable as a pattern of behavior. Meaning emerges for the observer.
Representation.
Need a representation system in the system memory.
There is a direct representation in the physical world. Everything is explicit because everything takes place in the physical world. It needs no explicit representations or models of the world. Only the physical world can ground the internal representation of the robotic agent.
Worlds.
The worlds used by the model are simplified, static and predefined.
The worlds used by the model are real, dynamic and sometimes unpredictable worlds.
Paradigm.
Functional (functional modules of information processing).
Activity or behavior based.
Intelligent behavior.
For a given situation, there is only one behavior: the one foreseen in the design. The centralized system is in charge of scanning the environment and selecting the most appropriate tasks, which are executed in parallel.
For a given situation, there are several possible behaviors or appropriate to the situation.
Fragility/robustness.
The system is fragile, as centralized control can be lost at any time.
The system is robust because behaviors act as barriers to other behaviors: one behavior cannot reach inside another behavior. The layered system is efficient, robust, redundant, more fault tolerant, flexible and adaptive.
Environment perception.
The centralized system is in charge of scanning the entrno.
Individual modules, which perform low-level activities, "sense" the environment frequently and are sufficient to detect changes in the environment.
Evolution toward autonomous intelligent robots.
Not ready to evolve toward autonomous intelligent robots.
It is ready to evolve towards modeling mobile, autonomous, cognitive humanoid-like robots (to facilitate interaction with humans), with flexible, adaptive behavior and learning capabilities.
The objections
Oren Etzioni [1993] retorts to Brooks that robotics is neither necessary nor sufficient as a foundation for AI. And that the software environment based on softbots is better than the physical world environment because it has the following advantages:
The softbots are complete intelligent agents, interacting in the software environment through commands and interpreting the responses. The sensors of a softbot are commands that supply the softbot with information about the external world. The effectors of a softbot are commands transmitted to the environment to change its state.
The softbots are the best testbed for AI. Software experiments facilitate systematic research because they are easier to run, control, debug and repeat than physical-level robotic experiments.
Replication of a oftbot is very simple. Replication of a robot is difficult and expensive.
software environments are best suited for the study of topics such as learning, planning with incomplete information, intelligent user interfaces, distributed AI, etc.
The cost, effort and expertise required to develop AI applications is low. Experimenting directly with physical robots is difficult and expensive (in time and money).
It simplifies applications, as it avoids having to consider many complex aspects of the physical world.
The softbots can operate in dynamic environments, as in the real world.
The design is simpler. In the reactive model you also have to make a design (layers, connection between perception and internal representation, etc.).
The evolutionary argument is debatable, because evolution is not a gradual process, since jumps occur (punctuated equilibrium theory). Darwin was a follower of the motto (attributed to Linnaeus) that "nature does not make jumps". The modern theory of punctuated equilibrium qualifies the theory of evolution in the morphological aspect by the sudden appearance of new species.
The mean time between failures is longer for a workstation supporting the software environment than for a mobile robot.
The development of softbots should be the foundation for the development of mobile robots.
Robotics and softbotics are complementary methodologies. There are AI research topics that have to be studied in the physical world. And there are topics that must be done in a software environment because of its flexibility and convenience: responding to error messages, cloning softbots on remote machines, modeling databases, controlling users, etc.
Robotic Languages
The issue of what should be the most appropriate robotic programming language in the traditional AI paradigm has been the subject of debate and discussion since the emergence of the first computer-controlled robotic systems. There has been no consensus, so there is no universal language, no hardware-independent language, and numerous particular languages, each with its own specific model or paradigm, which has led to a "robotic tower of Babel". Some of these languages have come close to the ideal, among them:
IRL (ndustrial Robot Language). High-level language for programming industrial robots, independent of the controller.
RoboScript, from the company Robotic Workspace Technologies (RWT). According to this company, RoboScript is the first industrial universal programming language. It is based on VBScript.
Strips (Stanford Research Institute Problem Solver). It is an automated plan generator. The same name was later used to refer to the formal language of the inputs to this plan generator.
ADL (Action Description Language) is an enhancement of Strips as a planner for robots.
XRCL (Extensible Robot Control Language). It is a language based on XML (a descriptive language for representing structured data using markup tags). It is also an environment that interprets the language. As a language, it is easy to use and has been proposed as a standard for code sharing between different robotic systems. It is based on general and reactive behavior. A behavior is a set of fuzzy (fuzzy logic) rules. The robot planning is a finite state machine (a set of nodes and arrows). A node is a state and is a set of behaviors. An arrow is a transition between states and is a set of rules (fuzzy or not) that instruct the control engine to move from one state to another. It contemplates coordination and communication between agents.
Current languages, in general, are complex, difficult to understand, debug and maintain. Sometimes they require a host language (host) as in the case of FROB, which requires Haskell.
However, the tasks of a robot can be easily specified at the abstract level such as picking up an object, depositing an object, moving to a certain position, etc. This abstractly described behavior implies a number of low-level behaviors: the level of interaction with the physical world based on the basic primitives or actions of the robot. If these primitives are defined by Cartesian geometry using absolute or relative coordinates, linear or circular motions, rotations, etc. and specific robot characteristics are hidden, then the robot's commands are independent of its specific kinematic configuration.
This has led to the conception of a system consisting of 3 layers, where there is a logical separation between the high-level abstract behavior (the planning layer) and the low-level execution (that of the robot commands). And in between, an intermediate layer between the high and low level layers, responsible for expanding the abstract goals into low level commands, ordering their execution and handling exceptions.
A robotic language must contemplate the two robotic paradigms and the following abstractions:
Agents. In the hierarchical paradigm there is only one agent. In the reactive and hybrid paradigms there can be multiple agents. An agent can be a single task.
Interaction between agents. Agents can interact with each other. For example, inhibit or activate an input or output of another agent, activate or deactivate an agent, stop or resume tasks, and so on.
Environment. This is the internal space (or memory) of the robot, reflecting the external world.
Continuous algorithms. These are algorithms whose input data are continuous, coming from sensors, and whose output is also continuous. The input signals (subsymbolic) are transformed into symbolic ones and produce a symbolic result that is translated into a subsymbolic output signal.
Events. These are discrete-time occurrences. They only have values at particular times (e.g., the robot hits an obstacle). They are situations or conditions that require an action. Conditions can be of many types: time conditions, predicate conditions, value conditions, task completion conditions, etc.
Exceptions. These are special events. For example, an algorithm may face inconsistent or null inputs.
MENTAL and Robotics
MENTAL is especially useful for applications based on softbots in the following aspects:
Paradigms.
The hierarchical and reactive paradigms are unified, since the robot behavior is modeled with the same language. The reactive model can be implemented only with simple rules. The hierarchical model is implemented with more sophisticated code. In both cases a representation, either simple or complex, is required. In addition, the MENTAL paradigm links with the robotic philosophy.
Abstraction.
The same language is used as in artificial intelligence and the development of applications is simplified and made more understandable.
Continuous algorithms. They are implemented by means of generic expressions.
Events. They are implemented as generic "condition → action" type rules. It contemplates events of all kinds. For example, the simple appearance of something in the environment, of a given expression or an expression belonging to a pattern.
Exceptions. They are treated as events, that is, with generic rules.
World model.
Provides a model of the common world internally and externally based on primary archetypes. The degrees of freedom are the archetypes. This model is the maximum possible approximation of mind, consciousness and life.
Environment.
The robot's inner (mental) environment is a symbolic/abstract reflection of the outer physical environment. The inner environment is the linking element between input and output (action). Communication between components, agents or tasks are realized through the environment.
Perception and action.
Perception (after preprocessing that converts analog signals into abstract symbolic structures) goes directly to the robot's internal environment. Actuation is the transfer of the symbols into commands to the effectors.
Representation.
Symbolic representation is necessary for communication with humans and with other machines. It has the advantage that the communication can be data or code.
Execution.
It can be sequential or parallel. Tasks can also be synchronized.
Learning.
You can change your code to implement learning. Learning can be done through meta-rules, rules that generate new rules or update existing rules.
Simulation of physical robots.
MENTAL vs. New AI
Brooks' "New AI" compared to MENTAL leads to the following thoughts:
The higher the level of abstraction, the more easily problems are solved. And the level of the new AI is the lowest possible: the physical level. In contrast, the MENTAL primitives are at the highest level of abstraction.
The complexity of traditional AI is due to the lack of the right language. With MENTAL everything becomes easier and simpler, working with primary archetypes, which have the maximum expressive and combinatorial power.
Brooks states that "The best model of the world is the world itself", but this world is superficial. MENTAL is the best model of the world because it is based on primary, deep archetypes common to both the physical and mental worlds.
Brooks is right when he tries to take intelligence directly through small modules. With MENTAL, intelligence is already present in everything, as everything is built with the same archetypes.
Brooks states that intelligence must be built from simple elements. But the true and supreme simplicity is to be found in the primary archetypes and to progressively increase the complexity with the level of detail through instances (manifestations) and combinations of the archetypes.
The best evolutionary model is not bottom-up, but top-down, that is, from the general to the particular. It is the spiral model, the model in imitation of nature. [see Appendix - Spiral Development].
MENTAL is the ultimate approximation to consciousness. Consciousness emerges from primary archetypes and their combinatorics (linguistics), not from independent low-level activities. In the new AI there is neither language nor semantics nor consciousness.
Addenda
Origin of the term "robot"
The term "robot" has its origin in the 1921 novel "R.U.R. Universal Robots Rossum", written by Karel Capek [2004]. In this work, the Czech word "robota" appears to designate an anthropomorphic machine, whose purpose was to replace human labor, making it more perfect, productive and cheaper. RUR is the name of the robot factory, founded by physiologist Rossum, located on a European island. The manufactured robots initially seem happy to work for humans, but one day they rebel and exterminate the human race.
The robots described in the book are not as we think of them today (metallic mechanical devices, anthropomorphic in appearance, but clearly differentiated from humans), but biological entities that mimic humans and can think for themselves. They are conceptually close to androids and even clones of humans.
The word "robota" is derived from the Czech "rab", meaning "serf", "slave", "compulsory labor" or "forced labor", and was translated into English as "robot". The term became popular with the dissemination of the novel in English. The Spanish translation appeared in 1962, published by Alianza Editorial. According to Karel Capek, it was his brother Josef who actually coined the term.
The word "Rossum" evokes the Czech word "rozum", meaning "reason", "wisdom", "intellect" or "common sense", an allusion that the robots could think for themselves.
The term "robotics" was coined by Isaac Asimov, who is also the author of the famous 3 laws of robotics, first appearing in the story "Runaround" (1942):
A robot cannot harm a human being or, by inaction, allow a human being to come to harm.
A robot must obey orders given by human beings, except if these orders conflict with the first law.
A robot must protect its own existence to the extent that this protection does not conflict with the first or second law.
Cog
Cog (short for "cognition") was a humanoid-like autonomous intelligent robot project created by Rodney Brooks and Lynn Andrea Stein [Brooks & Stein, 1994]. It was built in 1993 at the MIT Artificial Intelligence Laboratory. The project had two goals: 1) to build a prototype general-purpose autonomous robot; 2) to understand the process of human cognition. With 21 degrees of freedom, endowed with vision, hearing and speech, with the ability to move and manipulate objects. The goal was for the robot to learn to "think" through a cognitive learning process based on experience and social interaction. The project was discontinued in 2003.
Conscious-Robots.com
Also known as ConsBots.com, is an Internet portal dedicated to the scientific research of machine consciousness. In this field, the following terms are synonymous: artificial consciousness, synthetic consciousness and robotic consciousness.
Bibliography
Brooks, Rodney A.; Breazeal, Cynthia; Marjanovic, Matthew; Scassellati, Brian; Williamson, Matthew M. The Cog Project. Building a Humanoid Robot. Internet.
Brooks, Rodney, A. Achieving artificial intelligence through building robots. MIT AI Memo 899, 1986.
Brooks, Rodney, A. From earwigs to humans. Robotics and Autonomous Systems, vol. 20, Nos. 2-4, pp. 29-304, June 1997.
Brooks, Rodney, A. Cambrian Intelligence. The Early History of the New AI. The MIT Press, 1999. Disponible en Internet (Google books).
Brooks, Rodney, A. Cuerpos y máquinas. De los robots humanos a los hombres robot. Ediciones B, 2003.
Brooks, Rodney, A. Elephants Don´t Play Chess. Robotics and Autonomous Systems 6, pp. 3-15, 1990. Disponible en Internet.
Brooks, Rodney, A. New Approaches to Robotics. Science, New Series, 253 (5025), 1227-1232, 13 Sept. 1991. Disponible en Internet.
Brooks, Rodney, A. Intelligence without representation. Artificial Intelligence 47, 139-159, 1991. Disponible en Internet.
Brooks, Rodney, A. Intelligence without reason. Artificial Intelligence Memo No. 1293, April 1991. Disponible en Internet.
Brooks, Rodney A. A robust layered control system for a mobile robot. IEEE Journal of Robotics and Automation, 1 (RA-2), pp. 14-23, April 1986.
Brooks, Rodney A. Integrated systems based on behaviors. SIGART Bulletin, 2(4):46{50, 1991.
Brooks, Rodney A.; Stein, Lynn Andrea. Building Brains for Bodies. Autonomous Robots 1, 7-25, 1994. Disponible en Internet.
Capek, Karel. RUR. Robots Universales Rossum. Círculo de Lectores, 2004.
Capek, Karel. RUR: La fabrica del absoluto. Minotauro, 2003.
Etzioni, Oren. Intelligence without Robots (A Reply to Brooks). Artificial Intelligence 14 (4), pp. 7-13, Dec. 1993. Disponible en Internet.
Hager, Gregory; Pembeci, Izzet. A Comparative Review of Robot Programming Languages. Internet, 2001.
Müller, Vincent C. Is there a future for AI without representation? Minds and Machines, 17 (1), 101-115, 2006. Disponible en Internet.
