Research‎ > ‎Seminars‎ > ‎Seminars 2010‎ > ‎

2010.07.16 - Jan Peters - Towards Motor Skill Learning for Robotics

  • We want robots that are compliant and that we are capable of interacting with.
    • We need to develop efficient learning methods
    • How can androids learn basic human tasks
  • Put internal and external sensory information into "state" and forces into "control"
  • Stochastic policy to remove dimensionality problem and allow for exploration
  • Need two types of learning
    • Learn from teacher
      • Try and minimize distance of self action distribution to observed distribution
    • Learn from rewards
      • Reinforcement earning by reward weighted by previous actions (Peters & Schaal 2007)
  • To help with the learning they do the following:
    • Utilized motor primitives
    • Utilize context
  • Learn tasks in operation space
  • Motor primitives are defined in the dynamical systems representation, these are made up of discrete movements and rhythmic movements.
  • The task hyper-parameters are in theory those that are task invariant.
Comments