Linked arm in arm, they march up mountains and crawl through tunnels.
Each module in the crystalline robot system attaches to another using a key and lock mechanism. This two-dimensional system moves and alters its shape across the plane by expanding and contracting, but it cannot move vertically.
Researchers are developing shape-shifting robots that can climb obstacles, drop down cliffs and fit into tunnels. Small, individual modules link to form a system that can take a multitude of shapes to travel over varied terrain. Two distinctly different designs could allow military and first responder personnel to reach past obstructions into previously inaccessible areas while remaining at a safe distance.
Daniela Rus, associate professor, Electrical Engineering and Computer Science Department, Massachusetts Institute of Technology, Cambridge, and her research team have created two types of modules that make up the different robot systems. The first, called the crystalline robot system, is similar to structures made of crystals in that each module is identical. The second, called the molecular system, consists of modules that resemble molecules; two cubes are connected to each other by an arm, or bond.
Because of the differences in the designs, each system has its own style of movement. The crystalline system is considered two-dimensional, moving horizontally but not vertically. The modules expand and contract by a factor of two and attach and detach from each other, allowing the robot to change shape to move. A key and lock mechanism connects the modules to each other. Rus explains that the process is similar to holding and then releasing hands. When two modules are placed side by side, a locking mechanism turns and holds them together. The same mechanism is used to release them.
“If you want to move module A from one place to another, you don’t have to physically move module A,” Rus says. “You can perform some operations that pull module A into the body of the structure and propagate that change to pop out somewhere else in the structure.”
Unlike the crystalline robot, the molecular robot moves by changing the location of individual modules within the robot. “With this robot,” Rus explains, “if you want to move a module from location A to location B, you actually have to use the rotational degrees of freedom to allow this robot to shimmy on the surface of the structure all the way from point A to point B.”
Rotational degrees of freedom, in which the robot rotates parts of itself, enable it to move not only forward, backward and at right angles but also at various degrees in any direction, including up and down. Two parts of each module are connected by a right angle joint, similar to an elbow. The joint that connects the two cube-like structures in each module provides some degree of movement, while the other 10 sides of the cubes have connectors—similar to the crystalline modules—that allow the modules to lock with other modules. Two of the connectors, one on each cube, can rotate, giving the system four rotational degrees of freedom: one at each joint between the cube and the elbow and one at each rotating connector.
By employing this free-moving architecture, the robot can create a tower by stacking modules on top of one another until the tower is tall enough to climb and move over the surface of an obstacle instead of going around it. “Going from point A to point B becomes a simple problem—just beeline for the goal. If you find an obstacle, just climb on top of it,” Rus explains.
In simulation experiments, the robots also were able to split from a large group into smaller groups with the same underlying behavior. However, when the small groups reached a point where they needed to climb over an obstacle that was too large for them, they could not move forward. Rus and her team are working to design the robots to communicate with other groups, calling them to reunite to be large enough to get over the obstacle. The modules also could detect a hole in a structure and reposition themselves to fill in the hole.
In theory, the number of shapes the robots can create is infinite. Researchers have been able to change these mechanical amoebas from one shape into an entirely different shape. Rus uses the example of starting with one shape—such as a couch—and transforming it into an electronic dog. The new shape not only looks like a dog, but it also can move like one as well. Other possible shapes include robot arms, robot bodies, snakes, insects and multilegged or fixed-shape creatures.
A suite of distributive algorithms allows the robots to move and to self-reconfigure. Rus explains that the algorithms are like a library of task-specific moves that enable the robot to reconfigure itself as its environment changes. As long as the robot knows how to change shape to fit into a tunnel, it can fit into any tunnel, regardless of the tunnel’s shape, size and makeup.
Most of the distributed algorithms and control strategies were developed by hand, implemented into simulation and then tested through several different kinds of techniques to prove that they are correct. Rus admits that this takes a long time. “For my group and my research, what is important to me is that if I develop a new capability, then I want to know exactly when that capability can be used and when that capability is going to fail,” she says.
Another approach is to have an operator program a command into the system that tells the robot to change its shape or move. Rus and her team are now working on automating this process so the robots can learn how to do specific tasks themselves.
Rus proved that through distributive learning algorithms the robots could learn from scratch the locomotion rules that the researchers had developed manually. Without any programming, the robot system was given the goal to move forward. The system used the reinforcement learning approach in which modules tried various types of operations. After many iterations, the robots figured out all the same rules that were synthesized manually, she notes.
Researchers are now trying to expand these preliminary results into a more general approach to develop distributive controllers for any kind of task in a self-reconfiguring robot. “Our current results are fairly early and preliminary, but we were able to do this in a task-specific way for the locomotion task,” Rus says. Using distributive algorithms also will help to implement new kinds of tasks and give the robots more adaptations than they have at the moment with the sets of manually implemented, or fixed, behaviors, she adds.
Other future capabilities include equipping the robots with a small camera or with proximity sensors. Although the crystal robots have demonstrated the ability to move forward until they encounter an obstacle and then reverse direction, Rus admits that they have not done any specific experiments targeted at surveillance with the robots. “But I think we are at the right point in time to look into it,” she says.
Most of the experimentation has been done in simulation because the shapes and movements require more working modules than are available in the laboratory. Currently, Rus has 20 crystal modules and four molecule robots. The physical robots are for proof-of-concept experiments only.
“I think what stands between us and having a working system that will do something in the physical world is adequate funding,” Rus states. “An investment on the order of multimillions of dollars will probably help us deploy such a system in the next five years.” The National Science Foundation is funding the project, but the funding is in small amounts, which is only adequate to develop ideas and conduct laboratory experiments.
Rus would like to create more elaborate systems and perform additional experiments, and she hopes that a proposal the team has submitted to NASA will be the first step. “We would like to pack up modules really tightly in the belly of a ship and send them up in space. Once in place, they could self-release and self-assemble to assist in large tasks such as assembly and repair tasks to minimize the amount of time and projects that astronauts have to do in space,” she says.
The biggest problems, however, remain mechanical, namely, the power source and connections. Rus notes that improving the reliability and speed of the connection mechanism and creating more reliable hardware are key challenges. At the moment, if a connection fails, the entire system is dead, and recovery is not easy. One of Rus’ students at the Massachusetts Institute of Technology (MIT) has begun to address this issue and is developing a way for the robot system to compensate if one of the connections fails.
Another mechanical problem is the robots’ size, which is restricted to the 2-inch-cube size. The team would like to make the modules smaller, but each robot must house its own computer, electronics, motors and power source to make the whole system work. These requirements currently dictate the size.
The power source, four 3-volt lithium batteries per module, presents another concern. Each set of batteries lasts only a few hours during experiments. “When you have only a couple of hours of experimentation, you need to cut and re-solder wires and multiple batteries into each module,” Rus explains. “Logistically, this power problem is making our experimentation more complex than we would like it to be.”
One possible solution is using solar power in outdoor missions. Another potential solution is connecting one of the base-level modules to the power supply and building connectors that can share power. Using this approach, the robots would not only hold hands but also would have fast communication and power lines through the connectors. Although these features do not exist in the current prototype, the next version of the system will address these power issues.
Another challenge the researchers must address is the speed of the molecular robots; aligning the connections takes time. The laboratory versions of the crystalline robots move like an inchworm, so the modules compress and decompress and do not need to make and break connections often to move. However, the molecular robot needs to make and break connections. It takes about a minute for a module to move up and over a neighboring module. Rus is working to improve the speed, but it is not one of the primary goals of the project.
Rus and other MIT colleagues have several additional research ideas that they want to explore to make the technology more usable and practical. “I think that sometime in the future all we will have are the modules and the ability to get these modules to come together to form whatever is needed at the time: either to go on a surveillance mission or maybe to make a bridge, a bench or a big post to hang a camera. The versatility you get from these systems will truly out-power the conventional systems and become useful sometime down the line, but we are not there yet,” Rus says.