• Tidak ada hasil yang ditemukan

ARTIFICIAL INTELLIGENCE FOR ROBOTS

EXERCISES

2.2 ARTIFICIAL INTELLIGENCE FOR ROBOTS

What constitutes a robot, and how is it different from a machine? the answer to this question has changed over the past eight decades. Humanoid automatons, such as those in ‘Rossum’s Universal Robots’ and ‘The Metropolis’, were modelled on the human body but lacked smooth human functions and limb orientations and were low on emotional content1. With progress in industry and manufacturing and the automation of the production line and the car industry, the concept of a robot was more or less limited to the arm manipulator which engaged in repetitive ‘pick and place’ operations. Robots for mobile and other superior

1Of course, these were never made into robots rather ˇCapek’s actors reflected such features in their roles as robots.

31

with the environment. Therefore (1) robots performing repetitive jobs such as industrial robots and arm maniplators, (2) those which lack in a very clear mandate such as the Martian rovers, (3) those with human appearance as automatons or humanoid robots suitable to the domain of social robotics and (4) futuristic robots made by extending biology, such as androids and cyborgs, are all covered by this definition. The definition is not limited to the popularly accepted mechatronic design for robots which marries the mechanical to the electronic with a processing unit. However, an engineering point of view tends to constrain the definition and a robot is to have electronics, mechanical hardware and a processing unit. An actuation without the electronics and processing unit is more in the automaton domain and controlled by compressed springs, pneumatic valves and/or hydraulic control, probably such as the ones designed by Philon and in Japanese Karakuri puppets. Since, the definition is based on the ability of the robot to engage in real-world tasks, without external control, it is tempting to consider that the robot is truly ‘thinking’ as it is processing data from sensors to a processing unit to actuation much like the human brain; however this is merely executing a code piece and not ‘thinking’ per se.Chapter 9will try to focus on the

‘thinking’ faculty and sentient action in robots. Particular goals for the robot may either be specific, such as line following, light tracking or picking up empty coke cans, or a number of chores which converge towards a predetermined profile such as a military robot, a nurse robot, a domestic help robot or an office assistant robot are discussed inChapters 7 and8.

A more AI centric definition is given by Murphy,

“a mechanical creature which can function autonomously ”

The definition specifically mentions ‘creature’ and conforms to anthropomorphism and dovetails the works of Toda [322] and Wilson [356]. Implicitly it also suggests that autonomous functioning overlaps with intelligent behaviour.

As a working definition as per the scope of this chapter, a robot is an autonomous or semi-autonomous agency that undertakes jobs under direct human control, or it is partially autonomous but supervised and groomed by human supervision, or completely autonomously. In later chapters, one will find that this definition is not enough and as we progress towards newer realms of agent-based robotics, the definition will need to be modified.

Early ideas of Artificial Intelligence were suggested by Alan Turing in the late 1930s with hypothetical models which he called the automatic machine, which was later named the Turing machine. This was the bare bones of a central processing unit and helped to design computers in the post-war era. These early concepts were made into a fledgling discipline by the pioneering effort of McCarthy, Minsky, Newell and Simon.

Artificial Intelligence can be partitioned into the following seven subdivisions [248].

1. Knowledge representation. How does the robot represent the world? In human

33 context, for simple jobs such as locomotion, we tend to use maps or landmarks and resort to previous knowledge and experience. A robot does it using lasers or sonar, a table in the real world will reduce to an array of numbers corresponding to the intensity as perceived by the sensor. Since the onboard microprocessor if not very powerful, these methods approximate the dimensions and reduce objects to assortment of cubes, cuboids etc. much like a minecraft world.

2. Natural language. Language is unique since such unification of syntactic and semantic structures exists only for human beings and not in animals and is the definitive underpinning of our cultural and social systems. Noted linguist Noam Chomsky opines that language is at the interface of the two prominent cognitive processes: it is sensorimotor for externalisation and more intentional and deliberative for conceptual mental processing. To make robots understand and respond to human voice comes into play only in designing and developing more sophisticated robots which can closely interact with human society. Natural language processing libraries and chatbots have been very promising and are discussed inChapter 7. Voice based systems are still being explored into and Siri from Apple, Cortana from Microsoft and Google Now are promising results.

3. Learning.Robots are programmed with a number of task specific manoeuvre, but these are not exhaustive and to perform efficiently it must learn from experience.

Popular learning paradigms are cased based approach, artificial neural network, fuzzy logic and evolutionary methods. Nearly all state-of-the-art robot have a learning module.

4. Planning and problem solving.Making plans or algorithmic steps to accomplish a goal and solve the problems encountered in this process is inherent to AI agents, and is often a mark of their performance. For simple robots, planning is largely motion planning. However planning is also required for more interesting tasks such as, solving the Rubik’s cube, a game of chess, sliding tile puzzle, building a stack of blocks, making a schedule of daily chores etc.

5. Inference is to develop a conclusion from incomplete or inaccurate data sets. A robot often encounters inaccurate data from the sensors. In order to encounter this and prevent complete system shutdown the robot has to rely on inference, and ensures continuance of the processes.

6. Searchfor a robot usually means a search in the physical space — searching for an object or a goal point, but it can also mean a heuristic search where the robot is searching out solutions in an analytical manner.

7. Vision has become an integral part of robotics. For human beings, vision is unique compared to the other senses and is the trigger for most of our motor actions, the same is true for most of the animal world, so efforts to invent models of intelligence which can manipulate its local environment will have to address vision. Psychologists contend that vision enables our inner world, and nearly every consequence of our actions is simulated in our inner world prior to acting it out in the real world. Vision has had an important place in AI since early days with the pioneering research of Gibson [121] and later Marr [221]. Vision appeals unlike any of the other senses, and the overlap of ‘looking’ and ‘seeing’ seems as a deliberative process involving fast processing from our brain, but in more recent times, enactive models have established vision as an exploitative sensorimotor model.

FIGURE 2.1 Agent-world cycle, the agent and the world interact cyclically: agent acts on the world, the change in the world influences perceptions of the agent.

2.2.1 What is an ‘agent’ ?

The term ‘agent’ is interchangeably used with robots, programs, behaviours, animated characters etc. and can mean software as well as hardware implements. Russell and Norvig qualify an agent as an abstract entity which perceives its environment using sensors and acts upon that environment through effectors. Mobile robotics research is replete with the nomenclature for ‘autonomous agents’, autonomy is more often defined on context, or as per behaviour.

Autonomy loosely would mean that no other entity is required to feed its input nor is any required to keep it running. The robots can sense and act to fulfill given and implied goals in a dynamic environment, and they can go on working without any external intervention for substantially long periods of time. Franklin and Graesser [111] suggest the following definition for autonomous agents:

“An autonomous agent is a system situated within and a part of an environment that senses that environment and acts on it, over time, in pursuit of its own agenda and so as to effect what it senses in the future.”

This agent is a part of the environment, and thrives on interactivity as shown inFigure 2.1. This line of thought leads to a classification as suggested by Luck et al. [210] shown in Figure 2.2, however since agency is context-based, a strict categorisation into agents and non-agents would be superfluous if not just redundant. Every agent is situated, and is a part of the world. It can interact with the world and change the world and also its own perceptions. Caution should be taken to distinguish software agent from just any program.

As an illustration, a program which prints a line of text is not an agent, because it works on an input from a user. It doesn’t have any facility to interact with the environment (in the real world) or other programs (in the software world). This output would not effect later programs that are run and it runs once and stops lacking temporal continuity. While characters in a computer game, (viz. the ghosts in pac man) are agents as they have their own perception and consciously interact with the world (pac man 2D universe), every action from the player has consequences which enables an action from the ghosts, thus dynamically

35

FIGURE 2.2 Autonomous agent definition, adapted from Luck et al. [210]

changing the environment and, once run the game characters keep doing their job until the end of the game.

It is rewarding to have a quick jog through the various types of definitions that has been suggested for agents. One of the earliest definition is by Virdhagriswaran, with an eye on mobile agent technology;

“ The term agent is used to represent two orthogonal concepts. The first is the agent’s ability for autonomous execution. The second is the agent’s ability to perform domain oriented reasoning”

Russell and Norvig, acknowledged the binding between the sensing and acting;

“An autonomous agent is a system situated within and a part of an environment that senses that environment and acts on it, over time, in pursuit of its own agenda and so as to effect what it senses in the future.”

Maes, from a roboticist’s point of view, added the inherent pursuit of the agent towards a set of goal;

“Autonomous agents are computational systems that inhabit some complex dynamic environment, sense and act autonomously in this environment, and by doing so realize a set of goals or tasks for which they are designed”

Hayes-Roth’s defintion made agent interaction as an overlap of perception, action and reason;

“Intelligent agents continuously perform three functions: perception of dynamic conditions in the environment; action to affect conditions in the environment;

and reasoning to interpret perceptions, solve problems, draw inferences, and determine actions”

These definitions can be further extended, viz. (1) the ability to perceive information from the world as a cognitive agent, as most mobile robots are, (2) to work in unison with a large number of agents to lead to collective agency as in robot groups and swarms — swarm

simultaneously strive towards its goal of collecting uranium. Both Simon’s and Toda’s models provided extension to Walter’s experimental work and related physiological concepts as starvation and hunger threshold making the hypothetical animal negotiate trade-offs between attaining goals (collecting uranium) and survival (not running out of fungus and continuing its energetic wellbeing). Simon had concluded that in a treacherous and life threatening terrain, a randomly selected path will greatly jeopardise the chance of survival, however there exist hints in the environment which an organism must exploit for survival.

Toda used similar ideas and also equipped the artificial animal with multiple sensors, early ideas of incremental learning and adaptability. Behavioural economics approach, dynamical systems approach and evolutionary approach are some of the design principles probed in the early 1990s, these developments were instrumental in the ANIMAT approach and interconnected the principles of anthropomorphism, autonomy and sensory-motor for the development process. Here, I discuss Toda’s model and how it has motivated research in the last 50 odd years.

2.3.1 Toda’s model for fungus eaters

Toda’s fungus eaters were humanoid, artificial creatures mining uranium at planet Taros, α-Sapporo star system served as the first models for autonomous agents. These artificial creatures were supposed to have the means of collection, locomotion and ability for decision making by virtue of information perceived from the environment. Toda’s seminal research paper titled, “ The design of a fungus eater: a model of human behaviour in an unsophisticated environment”, blended Darwinian-like survival instincts with sensory-motor based stimulus-response with models of rationality with commercial goals to search, mine and store uranium ores. It also incorporated game theory and behavioural psychology to develop the fungus eater as an asymptote for human behaviour. Toda’s efforts were an extension of Simon’s model, where it was suggested that the ability of making rational choices is not accrued into the agent but rather governed by its environment. Toda’s approach did not subscribe to traditional psychological methods of observation, test, experiment and statistics rather Toda suggested the study of complete systems, rather than only isolated aspects like planning, memory, or decision making.

The project Solitary Fungus Eaters (SFE) is set in the year 2061 when these artificial animals are sent to an imaginary planet named Taros, in the α-Sapporo star system, for collecting uranium ores. The source of replenishment for these creatures is a type of fungus that is typical to that planet. These fungus eaters took instructions from the base command and there was no communication among themselves. The fungus eaters would roam around collecting uranium, unloading it in marked containers until they became inactive due to lack of fungi (starvation) or accident. Toda’s Taros is a planet with not much to its environment,

37 and it is littered with black and white pebbles with the much valued fungi being grey in colour. However, project SFE is seen to be of low efficiency and Toda’s seminal paper enlists suggestions to improve performance;

1. The fungus eaters are designed as wheeled humanoids, Toda [261, 322] suggests that the bodily form should be determined after the study of the terrain, gravity, climatic conditions, humidity, temperature, topography etc.

2. Since the primary job of the fungus eaters is to navigate the terrain and collect uranium and fungus at Taros which is a flat topology, this merits a wheeled model rather than legs.

3. Toda develops detailed designs for the visual sensors or eyes, which should be at the top of the body for maximum visibility.

4. Height of the fungus eaters should be optimal, too tall may not be cost effective and might be unstable, while too short will hinder the process of surveying the terrain effectively.

5. Addition of an olfactory sensor to smell out the fungus is also recommended.

6. Choice programs, for a fungus eater let E(ur) be the expected amount of uranium acquired following path r;Σruis the uranium picked up by the fungus eater; letv0

be the initial amount of fungus storage prior to engaging in the exploit of path r; let Σrv the fungus eaten along the path r and wr the best estimated amount of fungus to be consumed to finish the path,

E(ur) =Σru+f(v0, Σrv, wr) (2.1) where the form of function2 f will depend on the adaptability of the fungus eater to the terrain. Toda further suggests thatf will improve with experience, suggesting learning andmemorising techniques. Toda develops stochastic models of the above choice program and melds path planning with the physiological constraints of the fungus eaters.

7. After the modifications, the fungus eater is supposed to have three sensors at its disposal: geiger counter for uranium, olfaction for finding fungus and visual sensors for navigation. Toda suggests development of programs for coordination between these 3 streams of incoming information. These are early concepts of sensor integration.

8. Toda later extends his model to consider emotion, irrational behaviour, adding conditions for stopping an ongoing process — which would concur with ‘exceptions handling’ and the halting problem — and finally Toda introduces predators at planet Taros to simulate prey-predator like social psychological and game theoretic models.

Toda’s model served as a blueprint for an autonomous and self-sufficient agency with a set of goals, working over long periods of time without any external support and incrementally adapting to the environment. Much like Walter, Toda also believed that this approach will lead to models of human behaviour and cognition.

2.3.2 Design principles for autonomous AI agents

Toda’s model was broadly built on sensory-motor principles, and the fungus eater lacked the ability for appreciable cerebral activities, so however good uranium miners they may

2Toda uses a linear relation,f(v0+Σrvwr) in his model, while I have usedf(v0, Σrv, wr).

happens due to perception-action pairings. Situated cognition involves a continuous process, where perceptual information continues to flow through the sensors which leads to motor action which in turn changes the environment in task-relevant ways.

Walking, tightening a screw, switching on a light bulb are some examples of situated cognition. Wilson points out that large portions of human cognition takes place

‘off-line’, without any task relevant input and output, thus without any appreciable participation of the environment, and by definition is not situated. Creative thought processes such as writing a letter or scripting musical notes are examples.

2. Cognition is time-pressured.Embodied cognition is often accompanied with the terms ‘real time’ and ‘run time’. Since, embodied cognition is situated, it requires real-time response from the environment. The time pressure, if not met, leads to

‘representational bottleneck’ and instead of a continuously evolving response the system fails to build a full symbolic model. Behaviour-based approach is a remedy to such bottlenecks, and proceeds by generating situation-appropriate action on the fly by considering real-time situated action as the basis for cognitive activity, which appreciably lessens the time pressure. However, such models of situated cognition cannot be scaled up and therefore never lead to a model for human cognition.

3. The agency off-loads cognitive work onto the environment.Situated agency attempts to use the environment in strategic ways, by manipulating the environment to attend to the job at hand, rather than fully shaping up the system response to the concerned behaviour. For example, navigation with a compass exploits the magnetic alignment of the planet to enable finding the right direction. Similarly for the task of assembly, the pieces are arranged or used nearly in the order and spatial relationships of the desired finished product. This off-loading happens because there is usually a limit on the information processing, physical limits on attention and a limited working memory available to the agent. Concepts of psychology and behaviourism confirm this facet, as is seen later in the chapter.

4. The environment is part of the cognitive system.Cognition is not an activity of the mind but is distributed across the agency and the situation as they interact and is the result of continuous agent-environment interactions. Therefore, the situation and the situated agency are a single system. Similar ideas have been expressed by Uxekull, from a biological point of view, as will be discussed later.

5. Cognition is for action.Unlike cognition as per traditional AI, embodied cognition is always action oriented. Perception is dynamic, real time and occurs in tandem with motor action.

6. Off-line cognition is body-based.When not situated in the environment, in the decoupled agency, cognitive processes are driven by mental structures which are similar to simulations of sensory processing and motor control. The concept of the inner word is discussed in detail later in the context of conscious agency.