PhD Thesis Defense: Stuart O. Anderson
The Design of Control Architectures for Force-controlled Humanoids Performing Dynamic Tasks
|Stuart O. Anderson|
Carnegie Mellon University
May 22, 2013, 10:00 a.m., NSH 1305
This talk is about improving the process of designing controllers for humanoid robots. It describes tools we designed that enable us to iterate faster when experimenting with control systems that aggregate multiple model based sub-controllers.
Many model based humanoid controllers can be considered approximations to a fully general, but computationally intractable, optimal controller. Typically, simplified models are used to make the control problem tractable, but must be well chosen to capture the most important features of the problem domain. Because it is difficult to know apriori which simplified representations will perform best, the ability to experiment with different representations is an important feature of an effective controller design workflow.
Control designs based on simplified state representations typically combine controllers that solve specific sub-problems into a single aggregate controller. Common examples include the aggregation of trajectory generators with trajectory trackers, and balance controllers with end-effector force controllers. The primary contribution of this thesis is a method for aggregating sub-controllers that automates constructing models of the behavior of those sub-controllers. This method allows tight runtime coupling between sub-controllers, while limiting the need to adjust several sub-controllers to compensate for experimental changes made to any individual sub-controller. We call this combination of tightly coupled sub-controllers with automated model building Informed Priority Control.
The talk will describe the application of Informed Priority Control to humanoid balance tasks, including the bongo-board balance task, and show experimental results in both simulation and hardware environments. We produced stable controllers for simple balance tasks in both environments, but were only able to demonstrate robust controllers for the more difficult bongo-board task in simulation. We include an analysis of the instability of the bongo-board controller in hardware, focusing on the role of unmodeled sensor dynamics.
Jessica K. Hodgins, Chair
Christopher G. Atkeson
Stefan Schaal, University of Southern California