IPAB seminar - Professor Patrick Van der Smagt

Title: Deep learning in robotics

Abstract: The introduction of light-weight drive concepts, impedance control, and variable-impedance actuation has significantly increased the pace of development and applicability of robots. However, even today the integration of robots with their sensors in complex tasks remains a research topic, typically not being able to cope with real-world scenarios. Advances in machine learning in the last years is about to radically change this situation. The efficiency and generalising behaviour of deep learning now allows us to perform end-to-end learning on complex tasks. By combining deep learning with optimal control, reinforcement learning, and similar methods, actions can be swiftly learned using raw sensory data. These new methodologies no longer require explicit modelling of the movement in state space, configuration space, or task space. In my talk I will demonstrate the power of deep learning in the interpretation of complex sensory signals in a robotic setting, and show how to efficiently generalise between latent spaces and configuration or task spaces.