Learning sensorimotor control from experience and demonstration
An intelligent agent should be capable of performing useful actions based on sensory observations – a feat known as sensorimotor control. The notion of usefulness depends on a specific situation – for instance, in a driving scenario useful actions would get the vehicle to the destination point as fast as possible without crashing or violating traffic rules. I will talk about two general approaches to learning sensorimotor control – learning from experience and learning from demonstration – and about recent research projects in our lab in both of these directions. In one, we train an agent to navigate in three-dimensional environments based purely on its experience, without any human supervision. In another, we use imitation learning to train an agent to drive in busy urban environments and follow passenger’s commands.
Alexey Dosovitskiy received his MSc and PhD degrees in mathematics (functional analysis) from Moscow State University in 2009 and 2012 respectively. He spent 2013-2015 as a postdoctoral researcher at the Computer Vision Group of Prof. Thomas Brox at the University of Freiburg in Germany, with research focus on deep learning, specifically unsupervised learning, image generation with neural networks, motion and 3D structure estimation. Since May 2016 Alexey works on deep learning and sensorimotor control at Intel Visual Computing Lab led by Dr. Vladlen Koltun.
ACM Chapters Computer Science in Cars Symposium 2017