In just a few years, according to some predictions, homes and businesses will have access to personal robots capable of doing everything from fixing breakfast to performing routine maintenance tasks on office and factory equipment.
But before robots can begin buttering toast or lubing roller bearings, their owners will have to figure out a way of instructing the machines, preferably in simple language rather than with a series of complex programming commands. Maya Cakmak, a Georgia Tech robotics researcher, who recently authored a paper on the challenge of integrating robots into everyday life, says that active learning?giving robots more control over the information they receive?might prove to be the key to faster and easier robot training.
?Our research aims at allowing end-users to program new capabilities on their personal robots,? she says. ?Such robots can assist users in everyday tasks ranging from household chores to delivering items around an office or factory.?
Working at Georgia Tech’s Center for Robotics & Intelligent Machines (RIM), Cakmak used Simon, a humanoid robot developed in the lab of Andrea Thomaz, an assistant professor in the Georgia Tech’s School of Interactive Computing, as a testbed for active learning instruction. Thomaz and Cakmak programmed Simon to acquire new tasks by asking questions and then designed a pair of experiments to help them learn more about how robots can acquire task-related knowledge.
In the first experiment, Cakmak designated several human volunteers to play the role of an inquisitive robot seeking to learn a simple task by posing questions to a human instructor. She discovered that most participants asked ?feature questions??queries designed to elicit insight about a specific task (e.g. ?Can I pour the lubricant at any rate??).
In the next experiment, Cakmak asked human volunteers to teach Simon new tasks by answering the robot’s questions and then rating those questions on how “smart” they were. Again, feature questions were asked most often, with 72 percent of the volunteers rating these responses as the smartest questions.
Cakmak says that the experiments show that robots can acquire the ability to perform tasks by inquiring about specific features of the work rather than simply asking a human to approve a particular type of activity (?can I do it like this??) or imitating a human performing the task. She feels that her research could have a significant impact on robot manufacturers looking to bring multipurpose robots to a large user base. ?This research will allow manufacturers of highly capable general-purpose robots to reach a wider range of consumers, since the flexibility to program the robot will cover a much larger range of needs,? she says.
Real World Robots
Within the next decade, Cakmak expects to see robots helping people at home and at work in ways that are currently unimaginable. ?Users will gain the ability to program their robots to perform a wider range of tasks to assist them and to customize and personalize the robot’s capabilities to better fit the user’s preferences and the environment which they operate in,? she says.
?Our research unlocks the potential for how these robots can assist humans, by letting the users program the robots to meet their needs, rather than restricting the capabilities of the robot to what specialized engineers program on it.?
A paper documenting the research, “Designing Robot Learners that Ask Good Questions,” co-authored by Cakmak and Thomaz, was presented in March at the Seventh ACM/IEEE Conference on Human-Robot Interaction (HRI).Read More