El día 17 de diciembre impartirá un seminario D. Juan Cortés, investigador del LAAS-CNRS de Toulouse en el Salón de Actos del Edificio Ada Byron, campus Río Ebro.
Hora: 9:30, 17/12/2018
Title:
Cost-based motion planning: algorithms and applications to robotics and structural biology
Summary:
In robotics, motion planning algorithms have traditionally aimed at finding feasible, collision-free paths for a mobile system. However, beyond feasible solutions, in many applications it is important to compute good-quality paths with respect to a given cost criterion. When a cost function is defined on the configuration space of the system, motion planning becomes a pathfinding problem in a continuous cost-space. The cost function associated with robot configurations may be defined from the distance to obstacles in order to find high-clearance solution paths. It may also be related to controllability, to energy consumption, or to many other different criteria. In computational structural biology, where robotics-inspirited algorithms are applied to simulate molecular motions, the cost function is usually defined by the potential energy of the molecular system. Computing low energy paths in this context is important since they correspond to the most probable conformational transitions.
In this talk, I will present a variant of the popular RRT algorithm, called Transition-RRT (T-RRT), to compute good-quality paths in high dimensional continuous cost-spaces. The idea is to integrate a stochastic state-transition test, similarly to the Metropolis Monte Carlo method, which makes the exploration get focused on low-cost regions of the space. The algorithm involves a self-tuning mechanism that controls the difficulty of this transition test depending on the evolution of the exploration process, and which significantly contributes to the overall performance of the method. T-RRT is a simple and general algorithm that can take into account any type of continuous, smooth cost function defined on the configuration space. It has been successfully applied to diverse robot path-planning problems as well as structural biology problems. In the talk, I will also explain variants of the T-RRT algorithm to guarantee (asymptotic) convergence to the optimal solution in an any-time fashion. Finally, I will talk about the parallelization of the algorithm for applications to very-high-dimensional problems.
