Modeling and Planning with Human Impressions of Robots
Abstract
Complex technologies require their human operators to develop a mental model capturing their expectations of the device and how it will respond to inputs. Service robots, in particular, require human operators to undergo many hours of training before they become proficient at operating the robot. This process involves significant manual effort on the part of the human to adapt to therobot. In this project, the PIs investigate technologies that permit service robots to be aware of how they are perceived by humans. This kind of self-awareness about human mental models enables two new capabilities in robots: (1) to adapt themselves to meet human expectations, and (2) to set human expectations to match robot capabilities. The key issue is that humans makedecisions that affect the robot, such as delegating a task; for a good collaboration, a robot needs those decisions to be both well-informed and predictable. This project seeks to understand the psychology of human mental models and build computational mental models usable by robots to improve their performance in collaboration with people.This project seeks to expand our understanding of the psychology of human mental models of robots by studying the formation and updating of impressions of robots over the course of an interaction. The PIs investigate how mental models are influenced by the success or failure of actions the robot performs, by social cues like eye contact, and by the timing of actions. Durability of impressions is also a concern; once formed, does a mental model persist over time and subsequent observations, or does it remain malleable? If the robot takes actions with the intention of influencing human mental models in order to set expectations, how can the robot ensure that its actions will be accepted as believable? And how to impressions formed" about one robot performing one task generalize to other robots and other tasks? These questions willincorporate the well-known ""no""velty effect"" within human-robot interaction.In addition to understanding the science of mental models, the project aims to develo"p techniques to construct computational mental models usable by robots to describe and predict human decisions. The challenge is that the robot s observations are limited to the human s actions; all internal human mental state that determines those actions is unobservable (latent) to the robot. A computational model needs some structure to organize the latent state in order to predict human impressions of the robot. A remaining challenge then is to calibrate the mental model by inferring the latent factors that affect the human s perception of the robot s competence. This project employs a Bayesian, nonparametric model of Markovian decision making. Additionally, the model utilizes a hierarchical structure to represent generalization. At any moment in time, the robot s experiences with a person will inform the simplest model thatcorrectly predicts the observations made in response to robot actions. Consequently, the robot will be able to adjust its behavior quickly and without the need for extensive training data.If successful, the project will enable service robots to be more easily deployed into the field by reducing the learning curve to operate the robot. Robots will continually anticipate and monitor human decisions and interactively set expectations accordingly. This paradigm leverages human intuition and previous experience to calibrate and match expectations. Consequently, robots willbe more effective in the field, leading to simpler and lower-cost robot deployments.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 13, 2019
- Source ID
- N000141912299
Entities
People
- Ross Knepper
Organizations
- Cornell University
- Office of Naval Research
- United States Navy