Autonomous intelligent robots are the next generationSergio Alejandro Moriello
At least six research fields today give structure to advanced robotics: the one that relates the robot to its environment, behavioural, cognitive, epigenetic or developmental, evolutionary and biorobotic. It is a vast field of interdisciplinary study that relies on mechanical, electrical, electronic and computer engineering, as well as in the physical sciences, anatomy, psychology, biology, zoology and ethology, among others. The foundation of this research is the Corporative Cognitive Science and the New Artificial Intelligence. Its purpose: to illuminate intelligent and autonomous robots that reason, behave, evolve and act like people.
Autonomous intelligent robotics is an enormous field of multidisciplinary study, which relies essentially on engineering (mechanical, electrical, electronics and computer science) and sciences (physics, anatomy, psychology, biology, zoology, ethology, and so on). It refers to highly complex automatic systems that have an articulated mechanical structure - governed by an electronic control system - and characteristics of autonomy, reliability, versatility and mobility.
In essence, "autonomous intelligent robots" are dynamic systems consisting of an electronic controller coupled to a mechanical body. Thus, these machines need adequate sensory systems (to perceive the environment in which they operate), a precise adaptive mechanical structure (to have a certain physical ability of locomotion and manipulation), complex actuator systems (to execute the assigned tasks) and sophisticated control systems (to take corrective action when necessary). (Moriello, 2005: 172).
This approach addresses robots that are embedded in complex and often dynamically-changing environments (Mataric, 2002). It is based on two central ideas (Florian, 2003) (Muñoz Moreno, 2000) (Innocenti Badano, 2000): the robots:
"They are embodied," that is, they have a physical body able to experience their environment directly, where their actions give immediate feedback to their own perceptions
"They are situated", that is, they are immersed in an environment; Interact with the world, which influences - directly - their behaviour
Obviously, the complexity of the environment has a close relationship with the complexity of the control system. In fact, if the robot has to react quickly and intelligently in a dynamic and challenging environment, the control problem becomes very difficult. If the robot, however, does not need to respond quickly, the complexity required to develop the control is reduced.
In essence, "autonomous intelligent robots" are dynamic systems consisting of an electronic controller coupled to a mechanical body
Within this paradigm, there are several subparadigms: "behavioural robotics", "cognitive robotics", "epigenetic robotics", "evolutionary robotics" and "biomimetic robotics".
This approach employs the behavioural principle: robots generate behaviour only when stimulated; That is, they react to changes in their local environment (such as when someone accidentally touches a hot object). Here, the designer divides the tasks into numerous and different basic behaviours, each of which runs on a separate layer of the robot control system.
Typically, these modules (behaviours) can be to avoid obstacles, walking, getting up, and so on The intelligent functions of the system, such as perception, planning, modelling, learning, and so on emerge from the interaction between the different modules and the physical environment where the robot is immersed. The fully distributed control system is built incrementally, layer by layer, through trial and error, and each layer is responsible for only basic behaviour (Moriello, 2005, p 177/8).
Behaviour-based systems are able to react in real time, as they calculate actions directly from perceptions (through a set of action-situation correspondence rules). It is important to note that the number of layers increases with the complexity of the problem. Thus, a very complex task may be beyond the capacity of the designer (it is very difficult to define all the layers, their interrelationships and dependencies) (Pratihar, 2003).
Another drawback is that, because of the presence of various behaviours and their individual dynamics of interaction with the world, it is often difficult to say that a series of actions in particular has been the product of a particular behaviour. Sometimes several behaviours work simultaneously, or switch rapidly between each other.
Although they may reach the intelligence of insects, systems built on this approach will probably have limited abilities, since they have no internal representations (Dawson, 2002). In fact, this type of robot has a great difficulty in executing complex tasks and, in the simplest ones, the best solution, the optimal one, is not guaranteed.
This approach uses techniques from the field of Cognitive Sciences. It deals with the implementation of robots that perceive, reason and act in dynamic, unknown and unpredictable environments. Such robots must have very high-level cognitive functions that involve reasoning, for example, about goals, actions, time, cognitive states of other robots, when and what to perceive, learning from experience, and so on.
If robots are able to develop their own cognitive abilities by themselves, we should avoid programming them "by hand" for each conceivable task or contingency
For that, they must possess a symbolic and internal model of their local environment, and the sufficient logical reasoning capacity to make decisions and to execute the necessary tasks to reach its objectives. In short, this line of work deals with the implementation of cognitive characteristics in robots, such as perception, concept training, attention, learning, short and long term memory, and so on (Bogner, Maletic and Franklin, 2000).
If robots are able to develop their own cognitive abilities by themselves, we should avoid programming them "by hand" for each conceivable task or contingency (Kovács, 2004). Likewise, if robots are used to have human representations and reasoning mechanisms, human-machine interaction, as well as collaboration tasks, could be improved. However, high processing power is required (especially if the robot has numerous sensors and actuators) and a lot of memory (to represent state space).
Development robotics or epigenetics
This approach is characterized because it tries to implement systems of control of general purpose, through a prolonged process of development or autonomous self-organization. As a result of interaction with their environment, the robot is able to develop different -and increasingly complex- perceptual, cognitive and behavioural abilities.
It is an area of research that integrates developmental neuroscience, developmental psychology and situated robotics. Initially the system may be endowed with a small set of behaviours or innate knowledge, but -thanks to the experience gained- it is able to create more complex representations and actions. In short, it is a question of the machine autonomously developing the skills appropriate to a given particular environment through the different phases of its "autonomous mental development".
The difference between development robotics and epigenetic robotics - sometimes grouped under the heading of "ontogeneticrobotics" - is somewhat subtle, since it refers to the type of environment. In fact, while the former refers only to the physical environment, the latter also takes into account the social environment.
The term epigenetic (beyond genetics) was introduced -in psychology- by the Swiss psychologist Jean Piaget to designate their new field of study that emphasizes the senso-motor interaction of the person with the physical environment, instead of taking into account only genes. On the other hand, the Russian psychologist Lev Vygotsky complemented this idea with the importance of social interaction.
This approach applies the knowledge obtained from the natural sciences (biology and ethology) and the Artificial Life (neural networks, evolutionary techniques and dynamic systems) to real robots, so that they develop their own abilities in intimate interaction with the environment and without human intervention.
Through a fixed design, it is difficult to get a robot to adapt (self-organize) to a dynamic environment that evolves -often- through chaotic changes. From there, evolutionary robotics can provide an adequate solution to this problem, since the machine can automatically acquire new behaviours depending on the dynamic situations that arise in the environment where it is located.
Through the use of evolutionary techniques (genetic algorithms, genetic programming and evolutionary strategy), we can decide to evolve the control system or some characteristics of the robot body (morphology, sensors, actuators, and so on) or co-evolve both.
Through a fixed design, it is difficult to get a robot to adapt (self-organize) to a dynamic environment that evolves -often- through chaotic changes
In the same way, we can decide to physically evolve hardware (electronic circuits) or software (programmes or control rules). However, little has been done about evolutionary hardware (Fernández León, 2004) and, normally, what is done is to first evolve the controller in a computer simulation and, later on, it is transferred to the real robots. The robot controller typically consists of artificial neural networks, and the evolution consists of modifying the weights of the connections of this network.
At present, the main drawback of evolutionary control is its slow speed of convergence and the considerable amount of time that has to pass to effect the evolutionary process on a real robot (Pratihar, 2003). Also, it is not appropriate in solving problems of increasing complexity (Fernández León, 2004).
Biomimic robotics, biorobotic or biologically-inspired robotics
This approach deals with the design of robots that function as biological systems, hence they are based on the Natural Sciences (biology, zoology and ethology) and robotics. Since biological systems perform many complex processing tasks with maximum efficiency, they are a good reference for implementing artificial systems that perform tasks that living beings perform naturally (interpretation of sensory information, learning of movements, motor coordination, and so on).) (Ros, et al, 2002). Although we can obtain different degrees of "biological inspiration" (from a vague resemblance to an acceptable replica), the ultimate goal is to make machines and systems increasingly similar to the original (Dario, 2005)
The advantage of building bio-robots is that, because we can study all their internal processes, they can be contrasted with the different organs of the animal from which they are inspired. At present, scientists develop lobsters, flies, dogs, fish, snakes and robotic cockroaches, to emulate -to a greater or greater extent- the robust, flexible and adaptable behaviour of animals. However, few machines resemble their natural counterparts.
Replicating biology is not easy and it may be a long time before biomimetic robots can be made that are truly useful. Another problem -perhaps the main one- is that, although the different processes of many of these living things are well known, the difference to their human equivalents is an abysm. In fact, the way in which man perceives and acts is extremely more complex than a lobster does, to give an example.
(*) The article has previously been published in Trends 21. Megatrends, and has been authorized by the author for publication in TEJ