News/Events >> Jean-Claude Latombe

2005 Arthur Schoffstall Distinguished Lecture Series

2005 Arthur Schoffstall Distinguished Lecture Series
FIRST LECTURE:

Jean-Claude Latombe
Stanford University


Motion Planning with Probabilistic Roadmaps

For over 15 years, a major research theme in my group has been the development of random sampling schemes to create efficient motion planners. The main outcome of this research has been the Probabilistic Roadmap approach (PRM) to motion planning. Originally, this approach was intended to compute collision-free paths of robots with "many" degrees of freedom - at that time, 4 or more. But, over the years, successive improvements (as well as faster computers) made it possible to handle robotic systems with several dozen degrees of freedom operating in complex geometric environments. PRM was also extended to solve planning problems with motion constraints other than collision avoidance, for instance, visibility, equilibrium, contact, and kinodynamic constraints.
Concurrently, PRM has also been applied to non-robotics applications, e.g., for animating autonomous digital characters, designing product that can easily be assembled and serviced, testing whether architectural designs satisfy building codes, providing interactive tools to navigate in huge virtual reality models, planning complex surgical operations, and studying folding and binding molecular motions. This lecture will consist of three parts. First, I will review the PRM approach and various underlying techniques, especially sampling strategies. Then, I will discuss the probabilistic foundations of the approach and related theoretical results. In particular, I will argue that the main outcome of PRM is what its success tells us about motion planning problems, rather than the approach itself. Finally, I will discuss the recent application of PRM to legged robots navigating on steep irregular terrain - more specifically, rock-climbing robots. This application, which requires processing several thousand planning queries, many of which are not feasible, raises new issues associated with the fact that PRM is only probabilistically complete. This lecture is based on the work of many students, including Jerome Barraquand, Tsai-Yen Li, Yotto Koga, Lydia Kavraki, Rhea Tombropoulos, Amit Singh, David Hsu, James Kuffner, Gildardo Sanchez, Mitul Saha, Tim Bretl, and Kris Hauser.

 

April 11, 2005
4:00-5:00pm
Refreshments at 3:30pm
CII 4050

SECOND LECTURE:

Jean-Claude Latombe
Stanford University

 

Robotics Algorithms for the Study of Protein Structure and Motion

Proteins are the workhorses of all living organisms. They perform many vital functions, such as storage of energy, transmission of signals, transport of molecules, and defense against intruders. They are large molecules made up of hundreds to thousands of atoms. Each protein is a bonded sequence of smaller molecules (picked from a collection of twenty naturally occurring amino-acids) that determines its distinctive kinematics - a long kinematic backbone with small protruding side-chains. In robotics, this backbone would be called a highly redundant serial linkage. Key problems in structural biology include structure determination, conformation sampling, and motion simulation.
This lecture will discuss how algorithmic tools originally developed in robotics can be used to efficiently solve these problems, by exploiting the distinctive kinematics of a protein. In particular, I will consider the problem of completing partial models of protein structures resolved from electron-density maps produced by X-ray crystallography. I will describe software based on fast-inverse kinematics algorithms to recover the structure of protein fragments that yield fuzzy electron density maps. This work was done in collaboration with Dr. Henry van den Bedem and Ashley Deacon at the Joint Center for Structural Genomics (Stanford Linear Accelerator Center). Several protein structures obtained using this new software have recently been deposited in the Protein Data Bank.
I will also consider the problems of sampling protein conformations and simulating motion using a Monte Carlo approach. Here, the main computational bottleneck is to update the proximity relation among the protein atoms. I will describe a new method - the ChainTree method - that significantly speeds up Monte Carlo simulation runs without affecting their outcomes. ChainTree is based on hierarchical distance computation and collision detection techniques originally developed in robotics and computer graphics. Overall, the results presented in this lecture demonstrate that one can achieve major computational gains in computational biology by exploiting specific structural properties of molecules. These results are mostly based on the recent PhD of Itay Lotan.

 

April 12, 2005
4:00-5:00pm
Refreshments at 3:30pm
CII 4050

URL: http://ai.stanford.edu/~latombe

 

Home | People | Research | Publications | Contact | News/Events | Download | Links