Robotics research has shown to have a major role in the development of Artificial Intelligence (AI). While many fields in AI are concerned with optimizing certain specific problems (e.g. game playing), robotics is a multidisciplinary field that combines a lot of interesting AI research like machine learning and planning.
In this paper, I present a general survey of the robotics field. I will cover two main areas of robotics research:
Although Artificial Intelligence has many interesting subfields, one of the most interesting topics for me this term has been the discussions about intelligent agents, how they interact with the real world, how they learn and how they think. The philosophical foundations for AI discussed in [16, Chapter 26] was intriguing and interesting for me also. Robotics research tries to construct robots that fit very well into these discussions and this area is thus a logical and relevant choice for me to delve into.
The term robot was originally introduced to the public in a 1921 play called "Rossum's Universal Robots" [3]. The term was used to reference artificial (but organic) workers and the term has since changed its common meaning into the electro-mechanical systems that we usually think of when hearing the word robot.
There is no correct of incorrect definition of a robot, though. In some definitions, even a microwave oven is a robot whereas in others, the robot is defined as being able to interpret and interact with the environment, make decisions, and have some meaning and awareness of the world [4].
The most general definition of a robot would probably be an intelligent agent, to use a term from the course book [16]. What intelligent means in this context is then up for interpretation. One interpretation could be intelligent as in: It can pass the Turing test
or: It is able to make wise decisions on its own
. In my opinion, however, there is not one single definition that fits well just as there is no single definition of what a human is. It depends on the eye of the observer.
Although distinguishing robots between helpers and thinkers is very crude, I think it makes sense. To be more specific, the helping robot is a robot that solves one or more problems and solves them well, i.e. it does not know anything other than solving those problems. These problems are usually jobs that are undesirable or difficult for a human to do in some way.
The Roomba robot is an example of a helping robot. It is an automated vacuum cleaner [14], and although some people might enjoy vacuum cleaning, it is arguable that this is a chore that is quite undesirable for most people. Roomba does not know anything else but to vacuum clean but it is still classified as a robot and it still incorporates AI principles.
Another example of helping robots is presented in [11][13]. In this case, the examined robots are medical helpers. As an example, the articles describes a bypass operation on a human carried out by a robot called daVinci. The robot is guided by the surgeons but has the advantage of being more precise (no shaky hands) and lowering the time needed for the patient both to stay at the hospital and to recuperate.
The articles that describe the examples given in the last section pose the question: What is the future? The answers are similar. Both domestic and medical robots are converging towards more intelligent machines that interact with humans. The question is whether interaction equals thinking. It can be argued that robots that e.g. care for elderly people still just are helpers and not thinkers and that emotional responses are built-in rather than learned from experience.
Currently, the best example of a thinking robot is probably the research described in [12]. A robot named Gordon -- or Frankenbot as the writer calls it -- is controlled exclusively by living brain tissue from rats. The purpose of the project is understanding how memory is stored in the brain and possibly gain some knowledge useful for fighting Alzheimer's and Parkinson's diseases. Because the robot is actual driven by real brain tissue (although from a rat), it must be safe to call it a thinking robot.
An interesting issue arises when dealing with thinking robots, namely the questions of robot ethics. The science fiction writer Isaac Asimov made an early contribution to this: His three laws of robotics [4]. They state that robots should never harm humans or themselves. In [15], a lot of interesting ethical questions (but no answers) are given.
In this section, I will go over some specific robotics research. I will focus on robots that interact with humans, a branch of robotics often called human-robot interaction (HRI). A general article about this subject can be found in [10]. One of the pioneers in this field is Cynthia Breazeal [2] and I will survey (some) of her research contributions.
According to [8], Breazeal was among the first to define the term sociable robot in her Ph.D. thesis [5] from 2000. Around the same time, she was heavily involved in the construction of the Kismet robot. [9] is an early paper (1999) by Cynthia Breazeal and Brian Scassellati that describes the design and use of the Kismet robot which is arguably one of the first sociable robots.
The paper starts out with explaining what a robot needs for interacting socially with a human. The authors call this conveying intentionality. It means that the robot has to have beliefs, desires and intentions or at least convince the human that is has.
According to the authors, humans convey intentions in the form of e.g. posture and facial displays, and human infants slowly learn how to recognize and response to this, forming a theory of mind
for the infant. This learning process is the basis for the development of the Kismet robot. In other words, the robot is meant to mimic and learn like a human infant.
Four kinds of innate responses are identified with human infants: Affective, Exploratory, Protective and Regulatory responses. These responses are intended to communicate a kind of state of mind to a caregiver (i.e. a parent). They are (with my own interpretations), respectively: Feelings (love me
), curiosity/desire (show me that
), protection/concern (protect me
) and suitable environmental stimulation (I want more/less
). The interaction between an infant and a caregiver is without language but, according to the authors, still resembles dialogue very much.
The Kismet robot is based on the above four categories and is able to respond according to this basic theory. It is able to show various human emotions like sadness and surprise using fifteen different facial features and four different vision features. It has a perceptual (vision) system that searches for three different features: faces, motion and color. It further has a motivation system that consist of so-called drives. The drives are either social, stimulation or fatigue. Kismet is i.e. responsive to people, objects (e.g. toys) and the need for rest. It can show emotions based on three conditions with different values: Arousal, valence and stance.
Kismet behaves according to different levels of what the authors call cross exclusion groups. The behaviors on each level compete with each other which eventually leads to a specific action being executed, e.g. greet or look away.
The reason why I include all these specific details is to show how the initial abstract concepts, ideas and observations of human infants turn into a design for a robots behavior and perception system. Although no specific algorithms are presented, the flow of perceived data is still apparent. The paper ends with specific examples of human interaction with the Kismet robot and concludes what has been accomplished.
Other work using the Kismet include [6].
While the Kismet robot was an early example of a social robot, a more recently developed social robot is Leonardo. In [7] from 2006, Breazeal et al. describe the Leonardo robot. Leonardo is basically an updated, more advanced and life-like version of Kismet (obviously, the inventors are the same) and I will not go into detail with the architecture of Leonardo. In the article, a specific issue with social robots are discussed, using Leonardo as the test platform.
The paper specifically tries to address the problems that arise when human teachers interact with the social robot in a wrong
way, e.g. when the robot is shown ambiguous demonstrations. This is a problem whenever the robot encounters a human that is not familiar with the specific way that the robot learns. For a social robot to be deployed in an everyday setting, this is naturally a problem of big proportions. Late in the paper, the authors put it this way:
... how to handle and resolve "misconceptions" is a significant issue in developing collaborative systems capable of goal inference and plan recognition.The paper begins by motivating the problem (described above) and explaining how the Leonardo platform works. As in the case with the Kismet robot, the authors use (apparently recent) knowledge about human behavior to model the Leonardo system. Their hypothesis (derived from Simulation Theory) is that the brain has dual functionality, not only generating our own behavior but also predicting the behavior of other persons. This concept is directly applied to the Leonardo robot. It works by having the robot maintaining a full set of perceptions and beliefs of the human it is interacting with as well as its own.
To demonstrate the concept, a simple exercise of getting the robot to turn on four buttons is shown. The tricky part is, that the human can only see two of the buttons and Leonardo thus has to come to the conclusion that the state of mind of the human is different than its own and that the request from the human is incomplete. Even though the task is simple, it still shows that Leonardo is quite sophisticated.
Finally, I will briefly present some very recent (2008) work by Cynthia Breazeal and Cory Kidd. In [8], a weight loss coach robot is described. The robot itself is much less sophisticated than Kismet and Leonardo but is still designed as a sociable robot. In the paper, the relatively simple design of the robot is described and some results from a study with 45 participants are presented.
The results from the study are very interesting. The participants get started with a weight loss program using either the robot, only the robot software (i.e. no physical gestures) or paper. The authors first present four hypotheses about the study and then show that they are all true:
Even though the weight loss coach is described as a sociable robot, I would categorize it as a helping robot. But it is definitely on the verge between helping and thinking because it is interacting with humans and planning its actions.
In the previous sections, I have only scraped the surface of robotics. The field is huge, partly because the definition of a robot is pretty vague as described in section 2. Personally, I distinguish between helping robots and thinking robots. Research and development of helping robots is well underway both in household and the medical industry and new inventions and improvements are made all the time. Technologies like the weight loss coach and the daVinci surgeon robots are both very interesting and useful and it will be interesting to follow the development of even more sophisticated helpers.
One of the problems with the helping robots is, that they often still need guidance from a human. They are not totally autonomous. The Roomba only needs to be turned on but leaving it all by itself in the living room and expecting it to vacuum perfectly is utopic. The technology is improving though, and it is probably just a matter of time before the helping robots do their jobs close to perfect.
It is more difficult to say anything about the future of the thinking robot. The Kismet Cynthia Breazeal is one of persons that drives this kind of research forward. However, in [1], she is asked directly if she think robots will be able to think in the future, and her response is:
My interest is not how to make robots indistinguishable from humans, but rather how to make robots that are compatible and synergistic with people.This is an interesting statement. Breazeal's sociable robots are modeled and designed on knowledge from (among others) neuroscience and psychology. I would argue that those areas are indeed concerned with studying mental processes, i.e. thinking, and it thus seems curious that Breazeal is not interested in getting her robots to think.
But this seemingly self-contradictory point of view might just be another way of saying either robots will never think
or robots will never learn to think in our lifetime
. In the end of course, it always leads back to the question about how thinking is defined. Maybe the thinking robot will just spawn by itself someday out of the endless amount of research going on in the field. Or maybe there will actually never be a thinking robot.
I think, that we will see a lot more experiments like the Frankenbot
in the future. This approach of combining living tissue with computers is a fundamentally different approach than the purely electrical robot (like the sociable robot). While learning about our own brain may be one of the reasons for pursuing that kind of research, we might at the same time find the key to real intelligence. Not just artificial.