Thanks to decades of research in cognitive science and educational psychology, scientists know pretty well how people learn new concepts. Thus, researchers at MIT and Harvard University have worked together to apply well-established human conceptual research theories to the problem of human robot interaction.
They reviewed previous studies in which humans have tried to teach robots new behaviors. The researchers determined how these studies could incorporate elements of two complementary theories of cognitive science into their methods. They use examples from this work to show how theory can improve our understanding of robot behavior by making robot conceptual models faster, more accurate, and more flexible.
Serena Booth, a computer lab student at the Interactive Robotics Group, says that people who create more accurate mental models of robots often cooperate better, which puts them in difficult situations like manufacturing and healthcare. and robots work together. Intelligence. (CSAIL) and lead author of this article. “Whether we help people create concept robots, they will still do it. And this conceptual model may be wrong. This can put people in serious danger. It's important to do your best to give this person the best mental model they can build," Booth said.
Booth co-wrote this article with Harvard researchers along with his advisor Julie Shah, an aerospace professor at MIT and director of the Interactive Robotics Group. Elena Glassman, MNG '11, PhD '16, PhD '16, Associate Professor of Computer Science at the Harvard John A. Paulson School of Engineering and Applied Sciences, has a background in human-computer theory and interaction and is a consulting project leader. . . . Harvard co-authors also include graduate student Sanjana Sharma and research fellow Sara Chang. The research will be presented at the IEEE Human-Robot Interaction Conference.
Theoretical Approach The researchers analyzed 35 research papers on human-robot learning based on two main theories. "Analog transfer theory" suggests that people learn by analogy. When interacting with a new area or concept, they are implicitly looking for something familiar in order to understand the new object. The "Learning Variation Theory" argues that strategic variation can reveal concepts that people find difficult to grasp. It goes through a four-step process: iteration, collation, generalization, and transformation as people interact with new concepts. Many scientific works contain partial elements of the theory, but this is most likely a coincidence, Booth said. If the researchers had referred to these theories at the beginning of their work, they would have been able to design more efficient experiments.
For example, when researchers teach people how to interact with robots, they often show many examples of robots performing the same tasks. However, in order for humans to create accurate mental models of robots, mutation theory requires looking at different examples of robots performing tasks in different environments, as well as identifying where they make mistakes. “This is very rare in the human-robot interaction literature because it’s not intuitive,” Booth says, “but we also need to look at negative examples to understand what non-robots are.”
These cognitive science theories can also improve the design of physical robots. If a robotic arm looks like a human but moves differently from a human, it will be difficult for people to create accurate mental models of the robot, Booth explains. The analog transmission theory states that humans match what they know (a human hand) with a robotic hand, so if the movements don't match, people can get confused and have a hard time learning how to interact with the robot.
Expanded Explanations Booth and his collaborators are also exploring how human concept learning theories can improve explanations designed to help people gain confidence in new, unknown robots. “In terms of clarity, we have a very big problem with confirmation bias. In general, there is no standard for what a description is and how it should be used. As researchers, we often develop and offer explanatory methods that we like,” he said.
Instead, he suggests that researchers are using theories to teach people how to think about how to use explanations, often generated by robots, to formulate guidelines that people use to make decisions. By providing a curriculum that helps users understand what explanatory methods mean, when to use them, and where not to use them, Booth said, they can better understand robot behavior.
Based on our analysis, we made several recommendations on how we can improve research into human robot learning. On the one hand, they suggest that researchers include analog transfer theory and instruct people to make appropriate comparisons as they learn to work with new robots. Booth says the guide helps people use the right analogies so they don't get scared or confused by the robot's behavior. They also include positive and negative examples of robot behavior and educate users on how strategic changes in robot "policy" will ultimately affect robot behavior in strategically different environments, helping people perform better and learn faster. I suggest that you can help. The robot strategy is a mathematical function that determines the probability of each task that the robot can complete.
“We have been studying consumers for years, but we can see how useful and useless it is to show this person based on our own intuition. The next step is to more rigorously formulate the rationale for this work in a theory of human cognition,” Glasman said. Now that the first literature review using cognitive science theory has been completed, Booth plans to test his recommendations by reconstructing some of the experiments he examined and testing whether the theory actually improves human learning.
You must be logged in to post a comment.