Finding the balance between precision and openness is the special province of language, and scientists are finding new ways to harvest value from emerging languages that strike that balance in new arenas. An international standards team is improving manufacturing processes by helping computers better understand process commands. World Wide Web Consortium (W3C) members are making it easier to share data on the Web and infer relationships about data from different sources. Icosystem Corp. scientists are using a food language to create a food processor prototype for space missions. And HumanMarkup.org researchers are working toward translating human motions into a computer-operable language.

Developed by researchers at the National Institute of Standards and Technology and colleagues in France, Germany, Japan, and the UK, the new international standard ISO 18629 is a process specification language (PSL) that provides clear and concise definitions of information going from one system to another, says Steven Ray, NIST division chief of manufacturing systems integration. "A lot of computerized systems are now talking to each other around the world, and the risk of misunderstanding that information is much greater when computer systems are sharing information with other computer systems," Ray says. "Computers are basically very stupid, and have no common sense."

Automated systems that handle technical or business information to support a manufacturing process need a language that standardizes meaning through rigorous definitions. ISO 18629 does that by weighing in on the precision side of the language balance. The PSL is based on first-order logic, which considers statements to be either true or false. On top of first-order logic, situational calculus serves as the foundational theory. The team has built a fully axiomatized ontology. "It's an ontology whose definitions are taken down to the most primitive definition we can think of," Ray says.

One axiom in the PSL is the time point—an instantaneous, zero extent location on a 1D time line. This lets you consider the equality of two points that have the same value or distance between two time points, or the distance between two points that aren't equal—called a duration or interval. So, you can build temporal relationships on this axiom and clearly define terms, free of ambiguity.

For example, Ray points out that the word "before" is an ambiguous term. "Do you mean the second task starts immediately after the completion of the first task?" he asks. "Or does it start any time after the start of the first task, or is there a space of time between the completion of the first task and the second task?"

Without the standard, a computer system could mishandle the manufacturing process of painting and shipping a part if it didn't understand the difference between "painting before shipping" and "painting and letting the paint dry before packing and shipping." You can rigorously define "before" in this case as the entire process of painting and drying so that the system completes the task as desired.

The PSL helps machines talk to each other in a standard way by using Knowledge Interchange Format (KIF) statements instead of prose. The resulting common-logic statements are precise and computer interoperable. So, you can feed the language into a compiler or a reasoning system and guarantee statements' consistency. This industrial-strength standard "is a huge advantage when you are trying to capture the information in the design of an airplane," Ray says.

Basing the standard on first-order logic enables extremely comprehensive statements—highly expressive but not always computable in a finite amount of time. Ray says that having to capture the nuances of meaning and make definitions explicit at the expense of computability is "not necessarily a bad thing, because you're not going to be solving problems with it." The PSL's scope is the domain of discrete activities, which have a start and an end and can be combined into hierarchies or sequences or forked paths of parallel activities. He sees researchers using the PSL as a foundation for other systems or languages.

The language has been adopted into the Semantic Web services language (SWSL), which is designed to advertise and define services (such as booking or banking services) made available on the Semantic Web. NIST hopes to see the PSL embedded in the next version of the Unified Modeling Language. "Anywhere you have a business application that manipulates action," Ray says, "those [applications] can benefit from being grounded upon definitions and languages built on PSL."

Leaning on the other end of the balance is the Web Ontology Language (OWL), de-signed for use on the Semantic Web, which provides more computability than precise definitions. By labeling information on the Internet, OWL lets computer systems combine and query against data from disparate sources. Like the PSL, OWL is independent from prose and eases communication between computer systems. OWL is developed over other language standards. On top of its XML is RDF, which Eric Miller, lead of the W3C Semantic Web Activity, describes as a common model for representing data. On top of RDF is RDF Schema, and then OWL. "You could think of it as a relational model at a certain level that's been grafted into the very backbone of the Web," says Miller.

OWL has several dialects to meet the needs of various communities. Based on descriptive logic, the OWL DL dialect is very computable and less precise, more open. Allowing many possibilities emulates that characteristic of the Internet and enables data sharing from disparate sources and anticipating as-yet-unknown data. "It doesn't stipulate that there is only one concept of a person or one concept of a title; it rather provides the mechanisms for allowing different communities the ability to create these words, to create these terms," Miller says.

OWL provides a more descriptive way of defining resource classes, the constraints that you can put on those classes, and how those classes might relate in terms of equivalence and properties. "From an AI standpoint, this might be viewed as another way of representing ontologies, and [the] model theories that go along with such an ontology to license the inferencing that one might be able to deduce from various kinds of descriptions," Miller says.

OWL is built on a modular framework for expressing information that's independent of the applications that created the information. Relationships between labels can be inferred, and new relationships, or links, discovered. Miller compares OWL to a Lego set. The modular pieces are consistent and can be recombined to create new objects even when the pieces are packaged in different sets for different reasons, such as user interest, price point, and target audience. Machines can use the language to share and even create new data. The W3C's next step in development for the Semantic Web is to create a standard for business rules.

Eric Bonabeau, CEO of Icosystem, is working on a project for NASA that exploits two aspects of language: its abilities to describe and generate. Bonabeau explains the challenge as a problem of combining high-throughput screening with human evaluation. He wants to solve this problem through a hybrid of AI and human intelligence. Only human beings can evaluate uniquely human experiences, emotions, memories, and feelings, but machines are experts at exploring the space of millions of things. Bonabeau is using both approaches to generate the possibilities that might just keep astronauts on long missions from getting bored with the question, "What's for dinner?"

With the need to produce many combinations from the not-so-variable ingredients likely to be taken on long space missions, Bonabeau is designing a prototype machine that will use human feedback to guide a machine search for interesting combinations in the food space. Using a food language developed by Hervé This that describes textures and includes food-processing terms such as "whip," "fold," and "bake," Bonabeau is working on a machine that will create new combinations of food.

The food symbols in This's language configure the machine and direct the insertion of ingredients into it. The machine transforms that set of ingredients into a particular string of language symbols that also represent a sequence of commands for creating an end product. The human tastes and evaluates the end product, determining its "fitness"—good and interesting or bad and boring. The machine selects candidates for artificial evolution on the basis of the human evaluation and creates a second generation of food products, represented by the symbolic language. Those products are a result of either mutating or mating the fit candidates from the first generation. The process can continue until the machine produces a new and acceptable food.

"Anyone will be able to use the machine and explore the food space any way they want despite just having a very minimum of ingredients," Bonabeau says. He sees a day when consumers could use just such a machine, downloading recipes from the Internet and creating ever-new epicurean experiences.

Rex Brooks, executive director of HumanMarkup.org, is working on creating a standard that can capture motion and translate that into a computable language, Human Markup Language, to help promote fidelity of human communication in digital information systems. HumanML will establish a consistent vocabulary for describing communication based in semiology and transforming that formal description into terms that computers can use. Semiology postulates that communication components consist of signs, signals, and symbols and that those components combined with proxemics and haptics, physical environmental proximity, and touching can explain how and why some gestures have different meanings in different cultures.

Sorting out such meanings can help human communication, and 3D representations of motion can incorporate that knowledge to improve digital communication. The gestural ontologies will be combined with tokenizing motion capture—the process of capturing the physics of movement and equating a word with the movement.

"Developing a high-level language that will take the word 'run' and turn it into a behavior that can be translated by a computer into the 3D depiction of someone running requires building standards," Brooks says. "We're just not quite there yet."

Finding new meanings, creating new objects, linking inferences and data, and enhancing human-to-human, human-to-machine, and machine-to-machine communication are just some of the possibilities of leveraging new languages and their potentials for precision and openness. Like the advent of the Internet, this new evolution in networked and hybrid communication could transform our culture and business trends.

"All your back-end systems can go out there on the Web, connect with who they want, ask the questions they want, get the information they need. And, by having that information in the form they can digest, they can reason over it," says NIST's Ray. "That is not possible today."

Earlier this year, the Baltimore Veterans Administration Medical Center teamed up to further the research of MIT scientists in the field of robotic physiotherapy by establishing the Center of Excellence on Task-Oriented Exercise and Robotics in Neurological Diseases. Robotic physiotherapy uses AI to help stroke patients regain movement in their limbs.

Although robotic-physiotherapy research has been going on for some time, with other groups specifically working on its application to stroke victims, the team's principal investigators credit themselves not only with forging ahead in the field but also with pioneering it.

"I don't feel shy to claim paternity of the field," says Hermano Igo Krebs, colead scientist on the project and principal research scientist in mechanical engineering. "We recognized the societal impact of this potential transformation of physical medicine and rehabilitation practices. We did that at least five years ahead of any other group, including the ones that want to share paternity with us. By pushing MIT to patent the concept itself [in 1995], we made everybody else cognizant of its potential impact and rewards."

Indeed, for MIT researchers, imitation may be the highest form of flattery. "From the outset our machines have been interactive, meaning the machine changes what it does based on what the patient does," says principal investigator Neville Hogan, who holds appointments at MIT in mechanical engineering and in brain and cognitive sciences.

This, he says, has only recently begun to be incorporated into training with other systems. "Also, we insisted on the importance of low mechanical impedance, meaning the machines can 'get out of the way' of the patient when appropriate, thereby allowing them to express movement when they can," adds Hogan. He says many other groups have also adopted this approach.

Most recently, the team created a prototype for the rehabilitation of stroke patients' ankles. The Anklebot is the latest in a sequence of modules that provide robotic therapy to restore various body functions.

The robot, which fits around the leg in a brace, helps improve movement in paralyzed ankles, which are common in stroke victims. To grasp how the research team conceptualized and developed the Anklebot, consider its precursor, a robotic arm called the MIT-Manus.

In MIT-Manus therapy, the video screen prompts the patient to perform a series of arm exercises assisted only by a robotic arm brace (the MIT-Manus). If the patient can't move his or her arm at all, the MIT-Manus moves it; if the patient can move, the MIT-Manus provides guidance and assistance. Similar to the concept of resistance weight training, the more movement the subject initiates, the less help the robot provides.

After 16 years of research culminating in six clinical trials involving almost 300 stroke patients, the MIT-Manus approach has proven successful. In double-blind studies, patients in the robot-assisted group showed twice as much improvement as those in the control group. Patients in the experimental group received sensorimotor training, in which the robot assisted them as needed; those in the control group received sensory training, in which the robot provided visual feedback but didn't assist them.

"As far as we know, we were the first to use robots to deliver therapy in continuous physical contact with patients," says Hogan. "That is one reason why our patient history is so much larger than anyone else's."

Researchers hope to expand their success in working on arms and ankles to shoulders, elbows, wrists, and hands.

Still, the challenges have been substantial. According to Hogan, locomotion and balance have been the biggest hurdles. Solutions to improving the robots' locomotion seem to be slowly emerging. For balance, the team turned to an approach called impedance control, which controls the relationship between the robot's motion and the force it exerts.

"One of the enduring challenges of robotics is that if you attempt to use a force feedback loop to control force exerted on an object, you risk a pathological instability, even if the robot is stable. You use very low feedback gains," explains Hogan. "The problem arises from interaction between the mechanics and dynamics of the object and the dynamics of the robot's control system. We programmed the robot to mimic a spring or a shock absorber. Unlike motion or force, which may be profoundly affected by contact, the relation between force and motion is unaffected by contact."

In applying impedance control, the team programmed the robot to behave like a carefully chosen, nonlinear damped spring. The approach is novel in its application to humans as opposed to objects. "Others have shown the effectiveness of impedance control in such things as robotic assembly and complex or fragile parts," he says. "We've shown the effectiveness of impedance control for gentle manipulation of frail stroke patients."

As for whether the physiotherapeutic robots are exhibiting their own intelligence, Krebs says it depends on how you define "intelligence."

"All our robots, including the Anklebot, have two forms of intelligence," Krebs says. First, he explains, they employ performance-based control schemes that use ideas of human motor learning and neuroscience and modify the demand as well as the amount of guidance or assistance to the patient on the basis of his or her performance. "In this way," says Krebs, "we can track what the patient is capable of doing, and then continuously challenge the patient to improve by delivering assistance a notch below his or her performance."

Second, the devices have what Krebs calls mechanical intelligence. "Our robots were designed with mechanical properties close to the operationally desired ones, which afford us to use simple control rules that modify the robot impedance properties and which do not depend on the environment properties—in this case the patient's arm or leg."

According to Robert C. Richardson, lecturer in robotics at the University of Manchester's School of Computer Science, some of the fundamental aspects of robotic physiotherapy still need to be addressed and, therefore, the MIT approach is exciting and welcome.

Still, Richardson says, "Robotic physiotherapy devices cannot be considered to implement artificial intelligence in the traditional sense." He says to apply true AI techniques would require a significantly greater understanding of the connection between therapy, action, and brain function combined with the ability to monitor or predict these attributes.

As for whether using robots to help stroke patients regain movement has an inherent mind-body connection, Krebs says, "If one was to use a mind-body metaphor [in discussing the research], most AI roboticists would pay too much attention to the mind aspect while paying lip service to the body component."

MIT researchers believe that their research into robot-assisted physical therapy is paving the way for a new era marked by robotic gyms and robot-aided exercise and rehabilitation.

"A robotic gym is already a short-term technological reality," Krebs says. "If there will be a show-stopper it won't be technology, but rather the health-care system's reimbursement system."

"[MIT's work] is very important and exciting research," says Maja J. Matari , founding director of the University of Southern California's Center for Robotics and Embedded Systems and director of the university's Robotics Research Lab. "The robotics component of the work focuses on intricate issues of mechanical engineering and signal processing, rather than robot autonomy and control."

Another branch of robotics, socially assistive robotics (Matari 's own research area), presents a complementary approach to hands-on rehabilitation robotics. In this approach, autonomous intelligent robots interact socially with stroke patients to monitor, encourage, and coach them in the course of their recovery process entirely without physical contact. "The two areas of research present complementary and parallel directions in robotics—one focusing on hands-on mechanical systems and the other on hands-off intelligent interactive aides," says Matari .

Says Krebs, "We are at the cusp of a revolution in physical medicine, which will move this 'mature' industry into the 21st century. The only question is when."