PUB - Publikationen an der Universität Bielefeld

Modern robotic applications pose complex requirements with respect to the adaptation of actions regarding the variability in a given task. Reinforcement learning can optimize for changing conditions, but relearning from scratch is hardly feasible due to the high number of required rollouts. This work proposes a parameterized skill that generalizes to new actions for changing task parameters. The actions are encoded by a meta-learner that provides parameters for task-specific dynamic motion primitives. Experimental evaluation shows that the utilization of parameterized skills for initialization of the optimization process leads to a more effective incremental task learning. A proposed hybrid optimization method combines a fast coarse optimization on a manifold of policy parameters with a fine-grained parameter search in the unrestricted space of actions. It is shown that the developed algorithm reduces the number of required rollouts for adaptation to new task conditions. Further, this work presents a transfer learning approach for adaptation of learned skills to new situations. Application in illustrative toy scenarios, for a 10-DOF planar arm, a humanoid robot point reaching task and parameterized drumming on a pneumatic robot validate the approach. But parameterized skills that are applied on complex robotic systems pose further challenges: the dynamics of the robot and the interaction with the environment introduce model inaccuracies. In particular, high-level skill acquisition on highly compliant robotic systems such as pneumatically driven or soft actuators is hardly feasible. Since learning of the complete dynamics model is not feasible due to the high complexity, this thesis examines two alternative approaches: First, an improvement of the low-level control based on an equilibrium model of the robot. Utilization of an equilibrium model reduces the learning complexity and this thesis evaluates its applicability for control of pneumatic and industrial light-weight robots. Second, an extension of parameterized skills to generalize for forward signals of action primitives that result in an enhanced control quality of complex robotic systems. This thesis argues for a shift in the complexity of learning the full dynamics of the robot to a lower dimensional task-related learning problem. Due to the generalization in relation to the task variability, online learning for complex robots as well as complex scenarios becomes feasible. An experimental evaluation investigates the generalization capabilities of the proposed online learning system for robot motion generation. Evaluation is performed through simulation of a compliant 2-DOF arm and scalability to a complex robotic system is demonstrated for a pneumatically driven humanoid robot with 8-DOF.