The focus of my dissertation and postdoctoral research was the development of algorithms for learning robot motion control, the pair teacher demonstration and human feedback with machine learning techniques. Within my lab at NU and the RIC, I plan to continue within this research area, and also to expand into related topics that include a more rigorous formulation for human feedback and demonstration-based learning at multiple (high to low) control levels.

Brenna and iCub robot The overall approach of both my dissertation and postdoctoral work initially demonstrates a robot behavior, and then uses human feedback for further policy development. Learning from Demonstration (LfD) is a policy development technique in which the learner generalizes a policy from the example executions by a teacher. Machine learning techniques are then used to derive a policy, able to predict an action based on the current world state, from the resultant set of behavior examples [RAS 2009]; my work takes the particular approach of directly approximating the function mapping states to actions via regression techniques. Demonstration has many attractive features for both learner and teacher, including being an intuitive medium for knowledge transfer from a human to a robot that furthermore does not require robotics expertise. LfD does however have some potential limitations, such as dataset sparsity or poor correspondence between the teacher and learner, who may differ in sensing or motion capabilities. Overcoming potential dataset limitations is crucial for good performance, since LfD policies depend heavily on the quality of the demonstration data from which they are derived. My research uses corrective feedback from a human to overcome potential demonstration dataset limitations; in particular, to provide feedback that corrects poor policy predictions.

B. D. Argall, S. Chernova, M. Veloso, and B. Browning. A Survey of Robot Learning from Demonstration. Robotics and Autonomous Systems. 57(5): 469-483, 2009.   [pdf]

B. D. Argall. Continuing Robot Skill Learning after Demonstration with Human Feedback. In Proceedings of the International Conference SKILLS, Montpellier, France, December 2011.   [pdf]

Segway Policy corrections for mobile robot control.
Corrections Tactile feedback for policy refinement and reuse.