Robot skills adaptation of learned models with task constraints

• Autores: Daniel Hernández García
• Directores de la Tesis: Carlos Balaguer Bernaldo de Quirós (dir. tes.), Concepción Alicia Monje Micharet (dir. tes.)
• Lectura: En la Universidad Carlos III de Madrid ( España ) en 2014
• Idioma: inglés
• Tribunal Calificador de la Tesis: Pedro Lima (presid.), Alberto Jardón Huete (secret.), Manuel Ferre Pérez (voc.)
• Enlaces
  • Tesis en acceso abierto en: e-Archivo
• Resumen

  • A major goal in robotics research is to develop human-like robotic systems capable of interacting and collaborating with humans. The ultimate goal is for a robotic platform capable of performing, autonomously, in the unstructured scenario of human's natural environment. Humanoid robots must carry out any number of tasks which their human operators could reasonably expect from them during the normal development of a typical working day. Working alongside human means dealing with continuously changing environments and a huge variety of tasks, thus the robots should have the ability to continuously learn new skills and adapt their existing skills to new contexts. Therefore, humanoid robots need to display intelligent behaviour. Key attributes required to consider the behaviour of an agent as intelligent are the abilities to learn and acquire knowledge based on its experience, the capacity to understand or comprehend current relevant features in the environment, the capacity for reasoning and, also the ability to adapt. A framework for humanoid robots needs to provide a minimum degree of intelligence, that is, the ability to sense the environment, learn and, adapt its actions to perform successfully under a given set of circumstances. Humanoids must be provided with systems that allow them to continuously learn new skills, represent their skill's knowledge, and adapt their exísting skills to new contexts, as well as robustly reproducing new behaviours in a dynamic environment in order to cope with working in continuously changing environments and performing a huge variety of tasks. In our context a skill is defined as a motor trajectory motion learned by the agent, an acquired ability for the execution of a task. A robot skill is a complex action movement, reproducible when appropriate, and generalize to different contexts. Learning systems are required to acquire skills and develop task knowledge of how to act. Algorithms for learning and extracting important features of task actions are fundamental in order to build intelligent behaviours. The Imitation Learning approach formulates user-friendly methods by which a human user can teach a robot how to accomplish a given task, and generalize the demonstrated movements across a set of demonstrations. To learn the skills motion, a time independent model of the motion dynamics is estimated through a set of first order non-linear multivariate dynamical systems. We employ SEDS algorithm to learn a global dynamical estimate of the motion, through a set of first order non-linear multivariate dynamical systems in a statistical approach, as movement primitives. Despite the Imitation Learning approache's clear advantages, it would still be impractical for the human operator to teach the robot the skills for every needed task and for every foreseen situation, since the number of demonstrations the human must provide to the robot to generate a new model of a skill could turn it into a tiresome and time-consuming process; furthermore, it wouldn't be possible to cover every required task and every situation. Therefore, it is necessary to extend the classical Imitation Learning approach to learning a skill model in a way that allows the adaptation of a robot previously learned motion skills to new unseen contexts. The models of a skill are adapted to generate a new task by a merger, transition, combination or update operation over the given robot skill models. To reproduce a task adapted for an unseen context the robot must be given knowledge of the state of the environment and the constraints of the task. Using both, the already learned model of a skill, and the extracted constraints information of the current task, the model of the skill can be adapted to reproduce the task. The robotic systems must be able to store and later retrieve and use their knowledge of learned skills. The aim would be to have a knowledge base of the robot available skills for reproduction. The knowledge base needs to hold ah l necessary information for reproduction of the skills in the environment. Knowledge of the task would be distributed among the representation of objects, actions and events of the task and the state of the world. This work is centred on the major idea of future robotic systerns, more specifically humanoid robots, that are capable of interacting with humans in their homes, workplaces, and communities, providing support in several areas, and collaborating with humans in the same unstructured working environments. The aspiration is to have humanoid robots acting as robot companions and co-workers sharing the same space, tools, and activities. Our focus is on topics concerning the learning, representation, generation and adaptation, and reproduction of robot skills. In this work a framework is proposed for the learning, generation and adaptation of robot skill models for complying with task constraints. The proposed framework is meant to allow: an operator to teach and demonstrate to the robot the motion of a task skill it must reproduce; to build a knowledge base of the learned skills, allowing for their storage, classification and retrieval; to adapt and generate learned models of a skill, to new contexts, for compliance with the current task constraints.