Tutorial by Elmar Rückert at SoftNet 2018

Meet ROB Professor Elmar Rückert at SoftNet from October 14 to 18, 2018 in Nice, France.


T2. Neural and Probabilistic Learning Methods for Robotics and other Domains
Prof. Dr. Elmar Rückert, Institut für Robotik und Kognitive Systeme, University of Lübeck, Germany

In this tutorial, I discuss state of the art probabilistic and neural models that can be used to predict complex motions of humans or robots. The models can handle partial observable, missing data and are robust to sensor noise, which is demonstrated in challenging human postural control studies. In this experiment, a Gaussian mixture model was used to predict goal directed right arm motions solely from observing the motion of the trunk or left arm. The model can be also used for model validation, classification and movement  analyses and is as such interesting for a broad range of research approaches working with multi-modal motion data.

In the second part of my tutorial, I discuss how recurrent neural networks can be used for motion planning and obstacle avoidance. The model is based on the probabilistic inference framework and can be trained through reinforcement learning. It can be used to explain neural recordings of mental replays and pre-plays in rats during navigation tasks and provides a probabilistic theory for more complex cognitive reasoning tasks. The tutorial will conclude with extensions of this neural network that can be trained from millions of data samples which was exploited for learning dynamics models in robotics.