Recent progress in robotics and cognitive science has inspired a new generation of more versatile robots, so-called developmental robots. Many learning approaches for these robots are inspired by developmental processes and learning mechanisms observed in children. It is widely accepted that developmental robots must autonomously develop, acquire their skills, and cope with unforeseen challenges in unbounded environments through lifelong learning. Continuous online adaptation and intrinsically motivated learning are thus essential capabilities for these robots. However, the high sample-complexity of online learning and intrinsic motivation methods impedes the efficiency and practical feasibility of these methods for lifelong learning. Consequently, the majority of previous work has been demonstrated only in simulation. This thesis devises new methods and learning schemes to mitigate this problem and to permit direct online training on physical robots. A novel intrinsic motivation method is developed to drive the robot’s exploration to efficiently select what to learn. This method combines new knowledge-based and competence-based signals to increase sample-efficiency and to enable lifelong learning. While developmental robots typically acquire their skills through self-exploration, their autonomous development could be accelerated by additionally learning from humans. Yet there is hardly any research to integrate intrinsic motivation with learning from a teacher. The thesis therefore establishes a new learning scheme to integrate intrinsic motivation with learning from observation. The underlying exploration mechanism in the proposed learning schemes relies on Goal Babbling as a goal-directed method for learning direct inverse robot models online, from scratch, and in a learning while behaving fashion. Online learning of multiple solutions for redundant robots with this framework was missing. This thesis devises an incremental online associative network to enable simultaneous exploration and solution consolidation and establishes a new technique to stabilize the learning system. The proposed methods and learning schemes are demonstrated for acquiring reaching skills. Their efficiency, stability, and applicability are benchmarked in simulation and demonstrated on a physical 7-DoF Baxter robot arm.
Efficient and Stable Online Learning for Developmental Robots
Effizientes und stabiles online Lernen für "Developmental Robots"
2021
Miscellaneous
Electronic Resource
English
Developmental Robots - A New Paradigm
NTIS | 2005
|Learning Energy-Efficient Trotting for Legged Robots
Springer Verlag | 2022
|