Can we teach machines to learn from human demonstrations? This is the fundamental question that is driving Trevor’s research. As we demand greater performance from our robots in unstructured environments, it is becoming clear that the traditional, purely model-driven approach is not robust enough to handle complex tasks. Machine learning techniques have allowed us to create models from data that outperform their human-designed counterparts for isolated tasks – but designing algorithms that can learn robust, generalizable sensorimotor policies for mobile manipulators is still an open problem.
Trevor’s research involves attempting to learn, via expert demonstrations, the aspects of manipulation that are difficult or impractical to model by hand, while simultaneously using existing models to handle lower-level planning and control. He is also working to leverage unique sensor combinations, including force and contact sensing, which offer more information about the actions involved in completing a task.