Publications
The goal of this project is to investigate the interactive robot training paradigm for learning strategic and cognitive behaviors, i.e., using a human trainer to assist in the learning of such behaviors, assuming an unstructured and dynamic environment. In particular, we are studying behaviors that require (or at least benefit) from cultural bias or situational expertise that can be provided by a human domain expert. We have used the term interactive robot training (IRT) to distinguish it from previous Programming by Demonstration (PbD) work, as we are investigating other interaction methods besides just demonstration.
The motivation behind our research is a long-term vision for making robots easier to use in the real world. The vision focuses on the development of intelligent robot tools and assistants which operate with a human worker, making his job easier, extending his capabilities, and allowing him to avoid hazardous environments. To be convenient and flexible, fast interactive programming should be easily accomplished by the actual user. As a domain expert, the user knows best what specific task must be accomplished and the context in which it should be performed. As such, these domain experts provide the necessary situational expertise for purposeful, strategic tasks.
We assume that the environment is unstructured and dynamic. To be useful, our envisioned robot assistants should have the ability to perform skills safely and robustly in unstructured, uncalibrated, and unaltered environments. We assume also that the training may take place in an environment similar to the final execution environment, but not necessarily exactly the same. In some cases, it may be more convenient to train in a simulated environment but, again, not necessarily an exact duplicate of the final environment.
This work has presented some interesting challenges in human-robot interaction. A key element of our interaction method is the sensor-based qualitative state (QS) which serves as a link between the human trainer and the robot. We view the QS as a cluster in some sensor feature state, and use it in the skill model, as feedback to the robot, and as feedback to the trainer in describing the learned model. By abstracting sensory signals to a higher level, we allow both the human and robot to operate in their respective comfort zones. Some researchers have suggested that you must restrict sensory information to the human. In our work, we do not restrict it; instead we use the QS to guide the training process. In addition, we are building novel interfaces such as a wireless interface between a robot and a palm pilot device which can be used to direct the robot, display sensor state and learned skill model information, or even change the skill model directly.
These research issues are being addressed using a combination of computational intelligence techniques (i.e., genetic algorithms, neural networks, and fuzzy set theory). The project is coupled with the MU Computational Intelligence Laboratory. In particular, the project builds on expertise in pattern recognition, clustering, and spatial analysis.
skubic@cecs.missouri.edu
