Hau Tak Stanley Chan
PROJECT TOPIC: Research recent developments in the field of Artificial Intelligence, particularly concerning the COG project at the Massachusetts Institute of Technology (MIT). What is the purpose of the COG project? How does the COG project fit into the field of human-computer interaction?
Today, the computer is powerful in this world. It can help people work faster and handle complex problems. Most scientists do their research in computer technologies. Artificial Intelligence, AI, is the dream of many scientists. Scientists believe if they can succeed in developing AI, it will help people to do some dangerous jobs and assist handicapped people.
One supercomputer is named ASCI White. It is estimated to have a few hundredths of the processing power of a human brain. ASCI White was designed to perform tasks very different from those that a brain could perform. Its strength lies in its abilities to do math involved in simulating nuclear explosions. However, the supercomputer is completely unable to display human behaviors, such as learning and curiosity. One of the most advanced robots able to display a certain level of human-like intelligence is being developed in the COG Robot project at MIT (http://www.ai.mit.edu/projects/cog/). This robot is clearly not conscious, but it has been able to learn surprisingly life-like movements, such as fixating its eyes on a moving object, and reaching out its arms to touch objects.
Cog is not an acronym; Cog is a term that means “cognition.” It is a humanoid robot developed by the Artificial Intelligence Laboratory at MIT. This robot is used as a platform to bring together and explore the many subfields of human and artificial intelligence. The important parts are “Intelligence System.”
COG does not have brains like human beings; instead it is controlled by a heterogeneous network of many different processors. Each individual processor connects to others and is controlled by microcontrollers. The microcontroller is called Motorola 6811. These microcontrollers process inputs, such as responses from the strain gauges, perform servo calculations, and control motor responses. The audio and visual system is performing by Digital Signal Processor (DSP). The DSP relay data to the core processor network via ISA and PCI interface cards. This network consists of Texas Instruments C40’s. The core control network is run by a QNX real-time operating system. The operating system consists of 200MHz PCs running and connected to a motherboard by a 100VG Ethernet. Currently, the core network consists of four nodes; moreover, the number of nodes could easily be increased by plugging new nodes into the network.
The Sensory System is the second most important part of the AI COG project. Inside the Sensory System, there are four sub-systems that control vision and hearing. The four systems are Visual System, Auditory System, Vestibular System and Tactile System. The first of the four, Visual System, is similar to the human eye, which is binocular. Each eye can rotate in vertical and horizontal directions. Each eye is made up of two cameras inside, one for wide-angle view of peripheral visual field and the other for narrow-angle view of the center of the visual field. The Vestibular System consists of 3 semi-circular canals found in the human. They are implemented as 3-rate gyroscopes positioned on orthogonal axes. COG also has 2 linear accelerometers; all of them are mounted inside the head just below eye level. The Auditory System is enabled by two microphones, similar to those found in hearing aids. Stereo sampling of sound occurs at 22.05kHz with 8-bit resolutions. This system is heavily integrated with other modals such as vision, motor, etc. Sound localization has been achieved and current research is focusing on segregation of multiple sound streams. The Tactile system is also implemented. A 6x4 array of sensors in front of the robot torso can detect the position and force of a touch. Similar systems have been used on some of the hands built for robots.
A learning system is built in the robot. Generalization and abstraction of experiences allows human to learn and establish general rules to respond to different situations. It turns out that it is this learned mental activity that permits us to think and do so many things. We learn how to walk, do arithmetic, to write, to dress, to behave, and even to create intelligent learning systems. To a large extent, we learn how to think, to make plans, and to extrapolate. The learning process starts at birth and goes on intensively for many years until the end of schooling. Even then the learning does not stop.
According to my research, when an Artificial Intelligence System is written well, it will certainly think much faster than humans can. It could also be more precise than human thinking. Humans often get confused by concepts. Once an intelligence system in computers has been in use for many years, its concepts should be very precise and very detailed. We can easily copy and duplicate this structure of concepts and responses into an all-new intelligence system. So we see that in the long run, computers will think much better and much faster than humans ever can. However, a robot run by an intelligence system is still a machine. We give them objectives; they do not give us our objectives. Such computers will be of great help to humans in daily activities. We can keep them informed about all that is happening to us, through sensors that we carry, and they can give us their expert advice. This should make our choice of sub-objectives and our activities much more effective.
A robot is not easy to make. Inside a robot there are many interconnected parts. The main functions of the brain are sensing the environment, building a model of the present situation, choosing an action and learning from experience. The "Mind" is the activity within the brain. You are not your mind or your brain; you are the complete being, body and brain. From all the above, you will realize that to learn an activity, like driving a car or playing tennis, you have to reduce it to an automatic reaction to a given situation, namely to a response rule. Getting into the situation many times and repeating the response does this.
In the artificial intelligence world, most computer scientists work very hard. They focus on developing the brain control and sensory system. But those are technical problems. I felt the hard problem is the intelligence work on the robot. At the beginning, when I tried to start the paper I asked myself, “What is intelligence?” Some people say that any being that learns fast is very intelligent; some say that somebody who knows a lot must be very intelligent. But neither is exactly correct. People who learn extra fast or know a lot is of data are not intelligent if they cannot use what they have learned. To use their knowledge, people have to know when to use it. People have to sense which action or response is best suited for the environment or the situation at the moment. When people know about the situation they are in, then they can determine which part of their stored experience is appropriate for their present situation and do what is best for it. So putting intelligence into a robot is not easy, and we need a lot of time to develop these technologies. I hope our government can spend more money and equipment to support the computer scientists studying artificial intelligence.
Most people will ask why build a robot? Well, just imagine that you have two robots at home. There is something that you would like to do, like rearranging the living room. Naturally you will do that yourself, or sometimes with help from others. However, if you have robots, you may have the robots help you with some heavy items. Perhaps you don’t like to do housework like dishwashing, cooking, making up the bed and cleaning the house, you can also have your robot picking up these chores. With today’s technology, we still control robots by hand, since they do not have human knowledge. They have no ideas or mind of their own. Robots don’t know how to think. So if a robot has its own intelligence, it can immediately make quick decisions to choose a good way to finish the work or solve problems like humans. Robots cannot only help humans handle their housework, they can also help us save a lot of time and money. When you have a robot, you save money that you would otherwise pay to employees. The robot can be more efficient, so the working hours per week will shrink.
Today computer technology is very powerful. It is not easy to combine intelligence with computer. To create an intelligence system, we must have good equipment, good computer programmer, and computer knowledge. Inside an intelligence system, there are four sections. But I have only mentioned 2 sections in this paper. They are the sensory system and learning system. Those two systems are very important for an AI robot. And also those two systems are really hard to develop. So I hope in the future we can enjoy using robots and the robots can really help people to work and to solve problems.
Humanoid Robotics Groups. http://www.ai.mit.edu/projects/humanoid-robotics-group/
Robot Builder's Resources. http://www.robotics.com/resource.html
Return to the Top of this page
Return to the ESL 91 On The Web Home Page