Autonomous robots are tasked with producing stable and robust behavior even when facing unpredictable environments or changing conditions or in interaction with humans. In this talk, I will present an overview and outlook on my work in autonomous intelligent systems put into context of Artificial Intelligence.

Following a bottom-up approach, I will focus in this talk on three characteristics: First, decen-tralization as a central characteristic of biological motor control. Here, I will use locomotion in animals as a model for such adaptive systems that allows these to deal with varying conditions, e.g. when climbing or walking on uneven terrain. Adaptive behavior emerges from interaction of simple local control modules which allows the system to react quickly when facing disturb-ances.

Secondly, I will demonstrate how such a decentralized control structure can be leveraged in Deep Reinforcement Learning for much faster acquiring of more robust and general skills.

Third, this is extended towards a cognitive system using predictive internal neural network models: Cognitive behavior -understood as a form of planning ahead- leverages knowledge about performing a specific behavior into a novel context while minimizing any harm to the be-having system itself as it is using an internal simulation to predict possible outcomes. I will briefly present the underlying internal models and how these are grounded in lower levels as well as demonstrate how such predictive models can be leveraged for planning ahead in an internal simulation loop on the robotic system.