Goal Babbling

Both biological beings as well as robots need to deal with sensorimotor tasks that have many degrees of freedom. Goal babbling describes a concept to deal with this challenge and is inspired by newborns' way of learning to reach: already few days after birth they try to reach for goals, although they consistently fail. Mimicking these early goal-directed movements attempts turns out to be highly beneficial for machines as well! Read more ...

The Bionic Handling Assistant

The Bionic Handling Assistant is a pneumatically actuated robot that mimics an elephant trunk and is manufactored by Festo. It is very lightweight and passively compliant: just ideal for safe physical human-robot interaction. But: There are quite some challenges when one wants to utilize this potential. They span the entire range of simulation, control (answer: machine learning, of course), and software architecture. Read more ...

Neural Networks

Recurrent neural networks can compute any comptable function, but it's not possible just learn any network. A kind of networks that can be trained, even efficiently, are "reservoir" neural networks which only learn the very last layer. Turns out: they are really useful for all kinds of motor learning problems (e.g. humanoid whole body motion), because they can achieve very good generalization from very few training data. Read more ... (soon)

Cross-Modal Attention

One reason for human infants' incredible efficiency in learning is that they are guided and thaught by their parents. When parents explain things that are difficult for the infant, they need to guide the infant's attention and highlight important things. Therefore they, for instance, strongly synchronize gestures and speech, which is salient enough to be even detected on signal levels. So, what can computers and robots gain from this interaction pattern? Read more ... (soon)