CONTACT: Stanford University News Service (650) 723-2558

COMMENT: Prof. Oussama Khatib, Computer Science (415) 723-9753
e-mail: ok@flamingo.stanford.edu

Robot assistants: Mechanical brawn to help human brains

STANFORD -- Romeo and Juliet may not have much personality. And their look is decidedly industrial. But the pair of robots may be the progenitors of a new generation of robotic assistants who can help people with a wide range of domestic and physical chores.

"Our ultimate goal is to make robots that can act with complete autonomy. We can't do that yet, but we have reached the stage where robots can provide humans with semi-autonomous assistance," said Stanford roboticist Oussama Khatib.

According to Khatib, an associate professor of computer science, this means that robots can be designed and built so that they can cooperatively lift and carry loads in response to hand gestures from a human worker.

He and his students designed and built Romeo and Juliet last fall to demonstrate this point. The goal of the project, which is funded by the National Science Foundation, is to develop an automated mobile assistant able to move and position large or bulky objects under human supervision.

Each robot consists of a cylindrical base surmounted by a single, industrial-grade robot arm. It moves on a special set of wheels designed at Oak Ridge National Laboratory. The mobile robot builder Nomad Technologies helped in the robots' design and construction.

"Generally, robotics research has developed in two separate directions: mobility and manipulation. We are among the first to combine the two," Khatib said.

Romeo and Juliet can perform a variety of tasks that are simple for people but difficult for robots. These include washing windows, picking books off shelves, erasing whiteboards and dusting furniture. The two robots also have demonstrated the ability to cooperate on large tasks such as picking up and moving large objects that are difficult for an individual person, or robot, to handle.

More important, the two robots can be instructed to do these jobs by a relatively simple set of commands. Khatib has pioneered the development of a task-based approach to control that allows operators to specify the job that they want the robot to do and the robot then automatically figures out how to do it.

Today's industrial robots generally require a tremendous amount of programming to make them useful. Their controllers are not very sophisticated. As a result, they require detailed descriptions of the motion of every joint in the mechanism in order to execute a desired movement. When even minor changes are made to the task, the robots must be reprogrammed.

In contrast, task-level control allows the operator to instruct the robot to perform a task simply by describing the required motions of the robot's hand, or end-effector. The controller then generates the motions required by the rest of the robot to execute the task in a smooth, coordinated fashion.

With Romeo and Juliet, the researchers are pioneering efforts to integrate manipulation and mobility. To do so, they had to extend task-level control methods so that they work as well with a mobile robot as they do with a robot arm on a fixed base. In the process, the researchers found that the combined system could operate faster than for the arm alone. "It's much like throwing a baseball. You can throw it harder by using your whole body than you can if you just use your arm," Khatib said.

Combining a manipulator and mobile base into a single working unit means simultaneously addressing the limitations of the components. For example, present day industrial arms generally are fixed in place, in part because they require an environment that is precisely laid out and predictable. Such precise knowledge of the surroundings is much harder to maintain when the entire robot moves.

To address this problem, Khatib has developed controllers that rely much more on sensing the environment. Sensors provide the robot with rough equivalents to the senses of vision, hearing and touch. Distance measurements provided by vision and range sensors are used to control robot movement.

One aspect of this approach is called force control. This allows the robot to respond to forces acting on its hand. Sensors allow control of the pressure with which the robot grips objects or pushes against its environment. Thus, the robot does not need to know a book's precise size, shape, location and orientation to pick it up. Instead, the operator instructs the robot to move its hand to the book's approximate location, close its hand until it "feels" the book in its grasp, and then lift it.

Force control also makes tasks like washing a window pane much simpler. Just tell the robot to pick up a brush, dunk it into a pail of water, move it in the direction of the window and, when the brush pushes against the glass, begin scrubbing. This compares with directing the robot's hand to the location of the glass. If the robot is a fraction of an inch too close to the pane, the hand is likely to break it. Force control also makes it much easier for a robot to set an object down flush on a table, insert a bolt squarely in a hole or trace an irregular surface.

In the last year, the group made a major advance in the sensitivity of this type of control, allowing a person to guide the robot manually to desired positions simply by exerting a small amount of force on its hand or the load that it is carrying. A construction worker, for example, could use this control to guide a team of robot assistants carrying drywall to its final location.

The current generation of mobile robots require detailed descriptions of the world around them to navigate competently. This has made it difficult for robots to function in environments that contain unpredictable elements, like people.

"In the project, we are trying to deal with structures of control that use sensors. At the same time, we are trying to connect those with navigation, planning and intelligence," Khatib said.

The group has developed a method of route planning and obstacle avoidance called the "elastic band." Normally, a mobile robot carefully maps out a path it must follow before it starts moving. If something along its path moves, it must stop and make an entirely new plan.

With Khatib's approach, however, once a path has been determined, it is treated like an elastic band. The robot is programmed to feel an attractive force pulling it toward its goal. This gives the robot an initial direction. The robot also is programmed to feel a repulsive force pushing it away from obstacles. As obstacles move into its path, it is pushed away. The robot moves around the obstacle at a safe distance while continuing on its path toward the goal.

The researchers also have made important strides toward getting the two robots to work together. The first step was to reformulate the task and force control approaches: "This was a very difficult thing," Khatib said.

When two or more robots grasp an object, they temporarily become an "augmented object," in which the mass characteristics of the robots are added to those of the object. The researchers have found a way to determine the properties of this augmented object. As a result, they can apply task and force control in much the same fashion as they do in the single-arm case. The operator only needs to specify that the object be moved from one point to another and the controller will determine all the individual movements required to perform this action. This allows two robots to act together under centralized control.

In work first presented last fall, however, the researchers reported on a new approach that enables the two robots to work cooperatively on tasks without centralized control. In a centralized system, a master computer receives information from all the robots, synthesizes it, determines what actions each robot should take and orders them to do so. In Khatib's decentralized system, the robots produce their own plan of action based on information coming from their own sensors and from the other robot via a radio link. The researchers have used this approach to get Romeo and Juliet to pick up and move a long length of pipe.

Equipping robots with the kinds of general capabilities that Khatib's group has developed means they could be used for a variety of jobs: in an office building to deliver mail; in the home to act as a domestic aid; in hospitals to assist patients; in hazardous environments to handle dangerous materials; and in warehouses or construction sites to carry heavy loads.

In the last year, Khatib has seen a number of his theories validated by the performance of Romeo and Juliet: "We couldn't really check them without building the system. Well, we built it, and it worked!"


Oussama Khatib with robots.



Download this release and its related files.

The release is provided in Adobe Acrobat format. Any images shown in the release are provided at publishing quality. Additional images also may be provided. Complete credit and caption information is included.