ETL: Laying the Groundwork
for New Industrial Technologies

From robots that observe and cooperate to computerized "agents" that can visually identify and converse with individual humans, the Information Technology Department of the Tsukuba-based Electrotechnical Laboratory is conducting R&D projects that will help shape the future of computing.

by Steven Myers

Originally founded in 1891 as an electrical testing laboratory under the Ministry of Communications, Japan's Electrotechnical Laboratory (ETL) is world-renowned for its scientific achievements over the past century. These achievements include development of the world's first transistorized stored-program computer (1956) and the VLSI research initiative (of the 1970s). More recently, ETL has drawn attention for its development of the massively parallel IXM2 associative processor. The IXM2 can search 260,000 data elements in parallel, achieving the world's top performance for such AI (artificial intelligence) applications as real-time, speech-to-speech machine translation.

In 1979, the three ETL sites located around Tokyo were integrated and moved to Tsukuba Science City (in Ibaraki prefecture). Today, ETL is part of the Agency of Industrial Science and Technology (AIST), which is overseen by the Ministry of International Trade and Industry (MITI). Since 1988, the primary focus of the laboratory has been on conducting basic and exploratory research that could lead to new industrial technologies in the future. Like many other research institutions in Japan, ETL is notable for its long-term view and its willingness to commit resources to projects that may require many years before producing tangible results.

There are four major fields of R&D activity at ETL: information technology, energy technology, electronics and bioelectronics, and standards and measurement technology. ETL's Information Technology Department, the focus of this article, is divided into four separate divisions -- information science, computer science, machine understanding, and intelligent systems -- each of which is further subdivided into various sections.

Computing Japan's visit to the ETL Information Technology Department was arranged by Dr. Kazuhito Ohmaki, Director of the ETL Research Planning Office. The IT department is huge, currently hosting over 130 research scientists, so we were able to see only a small sampling of the myriad computer-related R&D projects underway.

As a government research institution, the atmosphere at ETL is naturally a bit more formal and bureaucratic than that at some other Japanese labs Computing Japan has visited, but the overall efficiency and organization of the IT department is impressive. Dr. Ohmaki arranged for presentations and meetings with the researchers in charge of three intriguing projects: a protocol mediation system already in wide use on the World Wide Web; a system of robots that can cooperate based on their "observations"; and an autonomous agent that can recognize individual faces and carry on conversations.

DeleGate -- Multipurpose
protocol mediation

DeleGate is a protocol mediation and conversion application, one that relays data through firewalls securely and efficiently, provides a cache that greatly improves response time, and performs both protocol and character code conversion. This highly practical system is the creation of Dr. Yutaka Sato (Computer Science Division, Information Base Section).

Dr. Sato began the project, in March 1994, to investigate and resolve problems related to security and character code diversity on the Internet. He made an early prototype version of the software available in June 1994 via anonymous FTP, and soon followed this with a version that is included on several CD-ROM collections of free/shareware software packages. DeleGate has quickly become very well known in Japan, and is currently in use at over 700 organizations around the world.

The basic functions of DeleGate are illustrated the figure, which depicts a typical situation of an organization employing a firewall to protect its internal data. The access control and routing control functions preserve the firewall security, while at the same time allowing for easy access and quick transport of non-sensitive data across the firewall. The protocol conversion function enables information packeted for Gopher, WAIS, or FTP protocols to be converted to HTTP for client presentations, and the character code function automatically detects and converts among the JIS, Shift-JIS, and EUC codes for Japanese characters. Finally, the
connection and data caches store recently accessed information to speed up response time if this same information is requested again.

Dr. Sato is currently at work on a number of enhancements to DeleGate, including the creation of an application-layer "virtual backbone" between DeleGate servers, internal functions for controlling cache size limits, expansion of the cache protocols to include FTP, NNTP and WAIS (the current version supports only HTTP and Gopher), and improvement of access control functions. He explains that DeleGate is nearing the end of the rapid-prototyping stage; he plans to continue researching the functions and structure of distributed information systems through further development of the DeleGate system.

Cooperation among robots
by "observation"

The researchers working on the Cooperation by Observation project have conceived a novel approach for achieving highly structured task coordination among autonomous robots -- an approach that does not resort to explicit communication or centralized control. Previous research in the area of robot cooperation has relied heavily on complicated planning methods and intensive inter-robot communication. Dr. Yasuo Kuniyoshi (Intelligent Systems Division, Autonomous Systems Section), leader of the six-member research group, explains that the emphasis of the ETL research is on developing an advanced capability for each robot to recognize other robots' actions through observation only, primarily using computer-vision methods.

The group has devised and implemented five basic algorithms to realize the visuo-motor functions necessary for the robots to coordinate their actions effectively. In order to test the effectiveness of these routines, they have developed a prototype system of four robots, each equipped with a stereo gaze platform and two onboard processors: one for controlling the mobile base and another for the gaze platform. Image processing is done on a remote host processor, which consists of pipeline image processors (Datacube Maxvideo system) and a CPU board running a real-time operating system.

Dr. Kuniyoshi presented demonstration video from the lab in which the ETL robots performed various tasks, with one fully autonomous robot acting as a helper for other robots running under fixed programs. The helper robot successfully performed such tasks as observing and removing obstacles from the paths of other robots. In another demonstration, one robot would push a can out to a certain point, then leave it to go fetch another can. The helper robot, observing the situation, would then go find the can and continue pushing it in the same direction as the first robot. The research group states that while there is still much to be done in order to create a versatile and flexible system of cooperating robots, the results so far have been extremely encouraging.

Multimodal agent-oriented interface systems

The Active Agent-Oriented Multimodal Interface project group at ETL (part of the Real World Computing Program) seeks to create a realistic human-like agent that can interact with users in a variety of ways -- not only through the recognition and production of speech, but also through visual recognition of the user. This project shares much in common with the Social Agent project underway at Sony Computer Science Laboratory. (See "Computer Science R&D in Japan" in the July 1995 Computing Japan.) Indeed, it is indicative of a general trend among many Japanese AI and computer science researchers toward the development of such agents as aids in human-computer interaction. Several university projects are currently exploring similar agent-related topics, though on a somewhat more limited scale.

Dr. Kazuyo Tanaka gave a demonstration of the ETL prototype system. He emphasized that one of the major differences between his group's system and those of other institutions is that the ETL system has successfully incorporated visual functions into the agent's communications process. When a user sits down at the computer, the agent appears, welcomes that person by name, and initiates a conversation.

According to the group, the system currently achieves 90% accuracy in identifying persons. The structure of the ETL agent system (whose architecture is shown in the figure) includes four separate subsystems for 3D facial display, vision, speech recognition, and speech synthesis. Communication among the subsystems is controlled by the interaction manager. (The 3D facial model used was created by Professor Hiroshi Harashima of the University of Tokyo).

The group plans several extensions to the system: the addition of hands to the agent, so that it can convey gestures; improvements to the image recognition system, so that individual persons can be more easily identified when they are in groups; and improvements to the speech recognition system, to increase the vocabulary and range of voices that the system is able to handle. The group also is carrying out a series of experiments with human subjects using the prototype system, in order to evaluate and improve upon the existing mathematical models for human-computer interaction.ç