In the latest phase of a project aimed at producing robotic primitives, or building blocks, that can combine to create complex, active forms, researchers here are tackling the problem of true three-dimensional self-assembly. If they are successful, a heap of seemingly inanimate objects may one day be able to morph into whatever shape is needed to perform the task at hand. More important, the lessons learned in programming such modular robots may simplify the integration of other large, multicomponent systems.

According to team member Mark Yim, the field of modular robots holds three promises: "versatility, which comes from reconfigurability and high degrees of freedom; robustness, from redundancy; and low cost, from batch fabrication." Still, he added, it is unclear whether the concept will succeed.

The advantages of building robots from many identical components seem compelling. Mass-producing tiny robotic blocks should be cheaper than designing, fabricating and building custom robots — especially if no one robot can be ideally equipped to do all required tasks. The combination of low cost and flexibility means that individual elements could be entirely disposable: Two broken robots, with the right working components, could easily transform themselves into one completely "healthy" machine.

Aside from robotics, the project could have other implications, which is why Xerox is interested. "Xerox' digital machines contain hundreds of individual components, many of which are computationally active," said Yim. "One of the hardest design tasks is programming these modules, which currently has to be done differently for every possible configuration."

Control software that could automatically adapt to different machine configurations "would be an enormous savings," he said. "At the lower end of our product lines, it would mean more options. At the upper end, it would mean smoother integration into complete solutions."

But the ideal system is not easy to design. For one thing, said Yim, "the number of possibilities grows exponentially with the number of modules. The key to developing these systems into useful machines is in the programming. Programming the distributed control of hundreds, thousands or millions of modules, each with their own processor and limited resources, is difficult, especially when their actions are tightly coupled."

There are other problems. "Redundancy does not necessarily imply robustness," said Yim. "More modules for redundancy also means more parts that can fail. So a proper architecture and control methodology are needed to exploit the redundancy to achieve graceful degradation."

In addition, "as the number of modules increases, batch fabrication may reduce individual module cost, but the whole system cost may still be high. We have yet to see what will happen here," Yim said.

The work started as Yim's PhD research in the early 1990s. At first, the blocks used were fairly limited in the ways they could be configured. Each module was essentially a cube, with two faces that were open, two that could bend (to provide actuation) and two that were fixed and could be used to join one module to another.

Later modules were similar, and though they could perform various types of locomotion, they worked best at forming simple shapes with essentially planar geometries. Yim said that a generation of that robot — which is now known as PolyBot — was the first ever to perform two fundamentally different modes of locomotion through reconfiguration.

Send in the robot

Yim sees PolyBots as most applicable to situations that could exploit its versatility — "things like search and rescue in a bombed or earthquake-damaged building." Here, the environment is highly unstructured so it is uncertain what shape of robot would be most appropriate. Planetary exploration, deep-sea mining, exploratory oil drilling, industrial or nuclear plant inspection, pipe maintenance in sewer systems and aid for the disabled are other possibilities. Clearly, the toy market could also be important. But repetitive jobs in defined environments would not be likely candidates. "While [PolyBots] may be good for a variety of applications, they will likely not be great at any specific one," said Yim. "For instance, if you knew that you needed a robot to paint a car in an assembly line, you could design a robot that does this quickly or efficiently. PolyBot may be able to do this, but it would most likely be less efficient."

The newer generation of robot, called Proteo, is still at the simulation stage. The building block here will not be a cube but a rhombic dodecahedron (RD), which Yim described as the three-dimensional equivalent of a hexagon. Such shapes can roll across each other, going from two faces in contact to the next two in contact, by executing just one kind of move: a 120 turn.

Once the Xerox team determined the shape and movement of its building blocks, it found that certain configurations that demanded large numbers of closely packed bricks were impossible to get into or get out of. That was solved by allowing the vertices of the RD blocks to bend.

Finally, the researchers had to figure out a way to allow the distributed intelligence of the modular robots to quickly create shapes specified by a controller. The result was a complex algorithm that defines the modules based on the desired shape.

Simulations have shown that many shapes can be "grown" successfully, though not all come out perfectly every time. The next step is to finish the design and fabrication of the module hardware.

The one-bot tool chest

Yim said that Proteo would be best for situations in which people aren't sure which tool they will need and can't afford to carry all that may help. The concept could be useful for larger jobs too, he said, since theoretically, Proteo "could form structures such as bridges and shelters automatically."

For the short term, Yim and his colleagues are looking at CAD/CAM. "The Proteo system can be used as a 3-D tactile user interface, which we call 'digital clay.' A user can manipulate the modules, and the 3-D representation of the system can be input into a computer system in real-time as well as give tactile feedback to the human."