Priming of relevant motor degrees of freedom to achieve rapid alignment of motor actions can be conceptualised as the rapid selection of low-dimensional action manifolds that capture the essential motor degrees of freedom. The present project investigates the construction of such manifolds from training data and how observed action trajectories can be decomposed into traversals of manifolds from a previously acquired repertoire. To this end we focus on manual actions of an anthropomorphic hand and combine Unsupervised Kernel Regression (UKR, a recent statistical learning method) with Competitive Layer Models (CLM, a recurrent neural network architecture) to solve the tasks of manifold construction and dynamic action segmentation.
Together with the advances in building anthropomorphic robot hands, we are facing the question of how to dexterously control such complex robots with up to 20 degrees of freedom in up to five fingers and a wrist. Implementing fixed grasp and manipulation programs does not lead to satisfying results. In our work, we propagate a manifold representation of such movements recorded from human demonstration. The main idea is to construct manifolds embedded in the finger joint angle space which represent the subspace of hand postures associated with a specific manipulation movement.
The spatio-temporal contact pattern during manipulation is a valuable source of information about object identity and object state, especially in uncertain environments. Using a bimanual robot manipulator setup with two 256 "pixel" touch sensor arrays, the present project is creating a "haptic pattern database" and investigates machine learning techniques to analyse the information contents of different haptic features and to extract identity and state information from haptic patterns. A closely connected goal are dynamic control strategies for contact movements with deformable or plastic objects, such as clay.
In the last decades, many researchers have used evolutionary algorithms to adapt the topology and connection weights of recurrent neural networks for various control tasks. This has become a useful machine learning technique. Because handling large genomes is difficult, however, these neural networks typically contain only a few neurons. If the genome contains a recipe for construction of the network instead of the network itself, it can be much smaller. We have developed a method than can exactly do this, and performs very well on a number of different problems.
Unsupervised Kernel Regression is a recent approach to learning non-linear continuous manifold representations. It has been introduced as unsupervised counterpart of the Nadaraya-Watson kernel regression estimator and uses this estimator to find both a latent space representation of a dataset and a smooth mapping from latent space back to the space of the original data. UKR has two main advantages: (1) one can apply leave-one-out cross-validation as an automatic complexity control without additional computational cost and (2) it requires only very few a priori specifications of parameters.
Unlike most existing approaches to the grasp selection task for anthropomorphic robot hands, this vision-based project aims for a solution, which does not depend on an a-priori known 3D shape of the object. Instead it uses a decomposition of the object view (obtained from mono or stereo cameras) into local, grasping-relevant shape primitives, whose optimal grasp type and approach direction are known or learned beforehand. Based on this decomposition a list of possible grasps can be generated and ordered according to the anticipated overall grasp quality.
The spatio-temporal contact pattern during manipulation is a valuable source of information about object identity and object state, especially in uncertain environments. Using a bimanual robot manipulator setup with two 256 "pixel" touch sensor arrays, the present project is creating a "haptic pattern database" and investigates machine learning techniques to analyse the information contents of different haptic features and to extract identity and state information from haptic patterns. A closely connected goal are dynamic control strategies for contact movements with deformable or plastic objects, such as clay.
In the last decades, many researchers have used evolutionary algorithms to adapt the topology and connection weights of recurrent neural networks for various control tasks. This has become a useful machine learning technique. Because handling large genomes is difficult, however, these neural networks typically contain only a few neurons. If the genome contains a recipe for construction of the network instead of the network itself, it can be much smaller. We have developed a method than can exactly do this, and performs very well on a number of different problems.
Unlike most existing approaches to the grasp selection task for anthropomorphic robot hands, this vision-based project aims for a solution, which does not depend on an a-priori known 3D shape of the object. Instead it uses a decomposition of the object view (obtained from mono or stereo cameras) into local, grasping-relevant shape primitives, whose optimal grasp type and approach direction are known or learned beforehand. Based on this decomposition a list of possible grasps can be generated and ordered according to the anticipated overall grasp quality.