Doctoral Seminar: Neural Networks as Add-on Modules for Improving Robot Performance
This is my Doctoral Seminar talk summarizing 5 years of her Ph.D. research in 20 minutes.
Bridging the Model-Reality with Lipschitz Network Adaptation
Model-based approaches have proven to be very efficient and effective when the dynamics is well-characterized. In this work, we explore how learning can be used to bridge the model-reality gap for uncertain systems to empower model-based design techniques. Related publication: RA-L 2021
Knowledge Transfer with Online Learning
A team of robots can often achieve more than single robots; however, training each individual robot in a team can sometimes be non-economical. If we are given a model trained on one robot, can we leverage the learned experience to enhance another robot? How do we characterize the similarity between robots, and what is the implication of having a higher similarity in the knowledge transfer problem? Related publication: ECC 2019
Quadrotor Impromptu Tracking with DNN
While it may be easy to design a controller to stabilize a system, tuning the controller parameters to achieve good tracking performance can be hard. Can we design a DNN add-on module to enhance the tracking performance of black-box control systems on arbitrary trajectories? Related publication: CDC 2017