Generative models for action generation and action understanding
Time: Thu 2019-09-12 13.00
Location: F3, Lindstedtsvägen 26, Stockholm (English)
Subject area: Computer Science
Doctoral student: Judith Bütepage , Robotik, perception och lärande, RPL
Opponent: Senior researcher Ivan Laptev, INRIA/ENS Willow team, Paris, France
Supervisor: Danica Kragic, Robotik, perception och lärande, RPL
The question of how to build intelligent machines raises the question of how to rep-resent the world to enable intelligent behavior. In nature, this representation relies onthe interplay between an organism’s sensory input and motor input. Action-perceptionloops allow many complex behaviors to arise naturally. In this work, we take these sen-sorimotor contingencies as an inspiration to build robot systems that can autonomouslyinteract with their environment and with humans. The goal is to pave the way for robotsystems that can learn motor control in an unsupervised fashion and relate their ownsensorimotor experience to observed human actions. By combining action generationand action understanding we hope to facilitate smooth and intuitive interaction betweenrobots and humans in shared work spaces.To model robot sensorimotor contingencies and human behavior we employ gen-erative models. Since generative models represent a joint distribution over relevantvariables, they are flexible enough to cover the range of tasks that we are tacklinghere. Generative models can represent variables that originate from multiple modali-ties, model temporal dynamics, incorporate latent variables and represent uncertaintyover any variable - all of which are features required to model sensorimotor contin-gencies. By using generative models, we can predict the temporal development of thevariables in the future, which is important for intelligent action selection.We present two lines of work. Firstly, we will focus on unsupervised learning ofmotor control with help of sensorimotor contingencies. Based on Gaussian Processforward models we demonstrate how the robot can execute goal-directed actions withthe help of planning techniques or reinforcement learning. Secondly, we present anumber of approaches to model human activity, ranging from pure unsupervised mo-tion prediction to including semantic action and affordance labels. Here we employdeep generative models, namely Variational Autoencoders, to model the 3D skeletalpose of humans over time and, if required, include semantic information. These twolines of work are then combined to implement physical human-robot interaction tasks.Our experiments focus on real-time applications, both when it comes to robot ex-periments and human activity modeling. Since many real-world scenarios do not haveaccess to high-end sensors, we require our models to cope with uncertainty. Additionalrequirements are data-efficient learning, because of the wear and tear of the robot andhuman involvement, online employability and operation under safety and complianceconstraints. We demonstrate how generative models of sensorimotor contingencies canhandle these requirements in our experiments satisfyingly.