Local learning for visual robotic systems
In this thesis a new supervised function approximation technique called Hierarchical Network of Locally Arranged Models is proposed to aid the development of learning-based visual robotic systems. In a coherent framework the new approach offers various means to create modular solutions to learning problems. It is possible to built up heterogeneous hierarchies so that different subnetworks can rely on different information sources. Modularity is realized by an automatic division of the input space of the target function into local regions where non-redundant models perform the demanded mapping into the output space. The goal is to replace one complex global model by a set of simple local ones. E.g. non-linear functions should be approximated with a number of simple linear models. The advantage of locality is the reduction of complexity: simple local models can more robustly be established and more easily be analyzed. Global validity is ensured by local specialization. The presented approach relies essentially on two new contributions: means to define the so-called domains of the local models (i.e. the region of their validity) and algorithms to split up the input space in order to achieve good approximation quality. The suggested models for the domains have different flexibility so that the local regions can have various shapes. Two learning algorithms are developed of which the offine version works on a fixed training set that is acquired before the application of the network, while the online version is useful if the network should be continually refined during operation. Both algorithms follow the strategy to place more local models at these regions of the input space where good approximation of the target function is harder to achieve. Furthermore, mechanisms are proposed that unify domains in order to simplify created networks, that define the degree of cooperation and competition between the different local models and that automatically detect data outliers to secure the application of a network. The value of the new approach is validated with public benchmark tests where several competitors are outperformed. The second major topic of this thesis is the application of the new machine learning technique in an adaptive robot vision system. The task is solved to train an arm robot to play a shape sorter puzzle where blocks have to be inserted into holes. To do so, different software modules are developed that realize interleaving perception-action cycles that drive the robot w.r.t. visual feedback. A visual servoing algorithm is presented that offers a simple principle to learn robot movements. It is based on the acquisition of training samples which represent observations of correct robot moves. The new approach to machine learning - specifically its features that are uncommon for supervised learning techniques - proves useful to realize this robot vision system. The possibility to combine different information sources in a hierarchy of local models helps to introduce application specific knowledge into the trained models. The outlier detection mechanism triggers error feedback within the system. The online learning algorithm makes the robot system robust against changes of its environment.