부산대학교 도서관

A central challenging problem in humanoid robotics is to plan and execute dynamic tasksin changing environments, and at the same time keep the result convincing and realistic.Sampling-based online motion planners are particularly powerful for automaticallygenerating collision-free motions in changing environments. However, without learningstrategies, each task still has to be planned from scratch, preventing these algorithms fromgetting closer to realtime performance. Moreover, the nature of the random samplingstrategy employed in these planners also results in extremely non human-like solutions.This document addresses these two issues by proposing to learn important features frompreviously planned solutions, or from real captured motion to improve both the efficiencyand the solution quality. Our methods work in changing environments, where obstaclescan have different positions in different tasks. However, we assume that obstacles arestatic during the execution of a single task. We first propose the Attractor Guided Planner(AGP), which extends existing motion planners in two simple but important ways. First,it extracts significant attractor points from successful paths as guiding landmarks for newsimilar tasks. Second, it relies on a task comparison metric to decide when previoussolutions should be reused to guide the planning of new tasks. The task comparisonmetric takes into account the task specification and as well environment features whichare relevant to the query.With combination of motion capture technique, the AGP planner also shows big improvementstowards generating realistic planned motions. We propose a constraint detectionmethod that applies to humanoid manipulation tasks. After recording a performer'sdemonstrated motion, our method will automatically detect important constraints, andthen segment the input motion according to different types of constraints. Attractors areplaced at the connections between each pair of segments and assigned the same constraintsas the previous segment. Then, given a new similar task, the new planning isguided not only toward the locations of the attractors, but also preserving the constraintsof the attractors.Several experiments are presented with different humanoid reaching examples whereobstacles are differently located for each task. Our results show that the AGP greatlyimproves both the planning time and solution quality, when comparing to traditionalsampling-based motion planners. We also show that with our constraint detection method,the AGP planner can efficiently find a solution that preserves the features of the input motion,making the solution motion coherent with the task being solved and therefore morerealistic. Although our current results are not yet capable of achieving real-time performancenor overall realistic humanlike motions, we believe that the techniques introducedhere are key for getting closer to these goals.