OATAO - Open Archive Toulouse Archive Ouverte Open Access Week

Motion Planning : from Digital Actors to Humanoid Robots

Esteves Jaramillo, Claudia Elvira. Motion Planning : from Digital Actors to Humanoid Robots. PhD, Institut National Polytechnique de Toulouse, 2007

(Document in English)

PDF - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader

Official URL: http://ethesis.inp-toulouse.fr/archive/00000454/


The goal of this work is to develop motion planning algorithms for human-like figures taking into account the geometry, kinematics and dynamics of the mechanism and its environment. By motion planning it is understood the ability to specify high-level directives and transform them into low-level instructions for the articulations of the human-like figure. This is usually done while considering obstacle avoidance within the environment. This results in one being able to express directives as “carry this plate from the table to the piano corner” and have them translate into a series of goals and constraints that result in the pertinent motions from the robot's articulations in such a way as to carry out the action while avoiding collisions with the obstacles in the room. Our algorithms are based on the observation that humans do not plan their exact motions when getting to a location. We roughly plan our direction and, as we advance, we execute the motions needed to get to the desired place. This has led us to design algorithms that: 1. Produce a rough collision free path that takes a simplified model of the mechanism to the desired location. 2. Use available controllers to generate a trajectory that assigns values to each of the mechanism's articulations to follow the path. 3. Modify iteratively these trajectories until all the geometric, kinematic and dynamic constraints of the problem are satisfied.Throughout this work, we apply this three-stage approach with the problem of generating motions for human-like figures that manipulate bulky objects while walking. In the process, several interesting problems and their solution are brought into focus. These problems are, three- imensional collision avoidance, two-hand object manipulation, cooperative manipulation among several characters or robots and the combination of different behaviors. The main contribution of this work is the modeling of the automatic generation of cooperative manipulation motions. This model considers the above difficulties, all in the context of bipedal walking mechanisms. Three principles inform the model: a functional decomposition of the mechanism's limbs, a model for cooperative manipulation and, a simplified model to represent the mechanism when generating the rough path. This work is mainly and above all, one of synthesis. We make use of available techniques for controlling locomotion of bipedal mechanisms (controllers), from the fields of computer graphics and robotics, and connect them to a novel motion planner. This motion planner is controller-agnostic, that is, it is able to produce collision-free motions with any controller, despite whatever errors introduced by the controller itself. Of course, the performance of our motion planner depends on the quality of the used controller. In this thesis, the motion planner, connected to different controllers, is used and tested in different mechanisms, both virtual and physical. This in the context of different research projects in France, Russia and Japan, where we have provided the motion planning framework to their controllers. Several papers in peer-reviewed international conferences have resulted from these collaborations. The present work compiles these results and provides a more comprehensive and detailed depiction of the system and its benefits, both when applied to different mechanisms and compared to alternative approaches

Item Type:PhD Thesis
Institution:Université de Toulouse > Institut National Polytechnique de Toulouse - Toulouse INP (FRANCE)
Laboratory name:
Research Director:
Laumond, Jean-Paul
Deposited On:21 Nov 2012 13:49

Repository Staff Only: item control page