Efficient Learning and Inference for High-dimensional Lagrangian Systems

Learning the nature of a physical system is a problem that presents many challenges and opportunities owing to the unique structure associated with such systems. Many physical systems of practical interest in engineering are high-dimensional, which prohibits the application of standard learning methods to such problems. This first part of this work proposes therefore to solve learning problems associated with physical systems by identifying their low-dimensional Lagrangian structure. Algorithms are given to learn this structure in the case that it is obscured by a change of coordinates. The associated inference problem corresponds to solving a high-dimensional minimum-cost path problem, which can be solved by exploiting the symmetry of the problem. These techniques are demonstrated via an application to learning from high-dimensional human motion capture data. The second part of this work is concerned with the application of these methods to high-dimensional motion planning. Algorithms are given to learn and exploit the struc- ture of holonomic motion planning problems effectively via spectral analysis and iterative dynamic programming, admitting solutions to problems of unprecedented dimension com- pared to known methods for optimal motion planning. The quality of solutions found is also demonstrated to be much superior in practice to those obtained via sampling-based planning and smoothing, in both simulated problems and experiments with a robot arm. This work therefore provides strong validation of the idea that learning low-dimensional structure is the key to future advances in this field.

Author dc:creator

• Vernaza, Paul N

Advisor dc:contributor.advisor

• Daniel D. Lee

Repository record dc:identifier.uri

https://repository.upenn.edu/handle/20.500.14332/29988

OAI identifier oai:identifier

oai:repository.upenn.edu:20.500.14332/29988