Date and Time:
August 7, 2023: 4:45 pm – 5:45 pm (JST)
Venue: Online and Open Space at the RIKEN AIP Nihonbashi office*
*The Open Space; AIP researchers are only available.
TITLE: Geometric Methods for Robot Learning: From Models to Representations
SPEAKER:Prof. Frank Chongwoo Park, Dept. of Mechanical Engineering and SNU AI Institute, Seoul National University
As robots become more high-dimensional and complex, and are asked to perform more challenging manipulation tasks in unstructured dynamic environments, the limits of traditional model-based robot planning and control are becoming more apparent. Efforts to augment traditional methods with, for example, models for friction, deformation, contact, external disturbances and noise, have had only limited success. Instead, there is growing optimism that by collecting large amounts of data — from a combination of trials and simulations — and applying machine learning methods — deep learning, reinforcement learning — traditional models can be bypassed entirely and replaced by a neural network. As of yet such optimism is premature; models and representations remain essential to effectively leveraging learning methods to real robotics problems.
In this talk we highlight three case studies in which methods from differential geometry and Lie groups play a central role in robot model and representation learning. First, exploiting a connection between the properties of a rigid body (link masses are positive, and inertia tensors are positive-definite) and the geometry of a certain Riemannian manifold, a set of robust algorithms are derived for estimating accurate models of humanoids and other complex high-dof robots, even when the measurements are noisy and incomplete. Second, methods for constructing accurate low-dimensional representations of a robot’s task-specific configuration space (latent space) are developed using the coordinate-free methods of harmonic mapping theory. Finally, a general method for constructing neural network models that are equivariant with respect to arbitrary symmetry groups — such neural networks are important in vision-based manipulation, for example — is presented.
Frank C. Park received the B.S. degree in EECS from MIT in 1985, and the Ph.D. degree in applied mathematics from Harvard in 1991. After joining the faculty of the University of California Irvine in 1991, since 1995 he has been professor of mechanical engineering at Seoul National University and member of the SNU AI Institute. He is a fellow of the IEEE and has held adjunct faculty positions with the NYU Courant Institute, the Interactive Computing Department at Georgia Tech, and the HKUST Robotics Institute in Hong Kong. His research interests span robot mechanics, planning and control, vision and image processing, mathematical data science, and related areas of applied mathematics. He is a former editor-in-chief of the IEEE Transactions on Robotics, developer of the EDX course Robot Mechanics and Control I-II, and coauthor (with Kevin Lynch) of the textbook Modern Robotics: Mechanics, Planning, and Control (Cambridge University Press, 2017). He is president of the IEEE Robotics and Automation Society for 2022–2023, and founder and CEO of Saige Research, an industrial AI company specializing in inspection and quality control.