[Thesis Proposal] Policies based on Trajectory Libraries

Martin Stolle
Robotics Institute
Carnegie Mellon University

Place and Time
NSH 3305
10:00 AM

Abstract
I present a control approach that uses a library of trajectories to establish a global control law or policy. This is an alternative to methods for finding global policies based on value functions using dynamic programming and also to using plans based on a single desired trajectory. Our method has the advantage of providing reasonable policies much faster than dynamic programming can provide an initial policy. It also has the advantage of providing more robust and global policies than following a single desired trajectory. Trajectory libraries can be created for robots with many more degrees of freedom than what dynamic programming can be applied to as well as for robots with dynamic model discontinuities. Results are shown for the “Labyrinth” marble maze and the Little Dog quadruped robot. The marble maze is a difficult task which requires both fast control as well as planning ahead. In the Little Dog terrain, a quadruped robot has to navigate quickly across small-scale rough terrain. In our past work, I have used global state to represent the knowledge in the trajectory libraries. In order to broaden the use of a library, I propose the use of local state representations, which allow the knowledge represented by a library to be used in novel situations. Three different mechanisms for this transfer are proposed: Information about the goal of a task can be explicitly represented in the local state. Libraries using this representation can be transferred directly to new tasks. Alternatively, the local state representation might not include a goal feature. When using such a library, a search over actions in the library has to be used to pick actions that obtain the goal. Finally, one can cluster the actions in the library in order to create abstract actions. This will simplify the search process.

Further Details
A copy of the thesis proposal document can be found at http://gs3020.sp.cs.cmu.edu/~mstoll/proposal.pdf.

Thesis Committee
Christopher Atkeson, Chair
James Kuffner
Drew Bagnell
Riger Dillmann, University of Karlsruhe