Department for Automation, Biocybernetics and Robotics





Kendama playing robot

Research activity duration: 2009 - 2010
Research area: Laboratory of Humanoid and Cognitive Robotics
Activity funding: Research Agency
Activity leader: Bojan Nemec
Coworkers:
Aleš Ude
Rok Vuga

Abstract

Supervised and Reinforcement learning of Kendama playing robot

Activity description

This project deals with robot playing ballin-a-cup game, also known as Kendama, Balero or Bilboqet game. The aim of this game is to swing the ball hanging down on a string from a cup and catch it with the cup.We evaluate two learning methods applied to the ball-in-a-cup game. The first approach is based on imitation learning. The captured trajectory was encoded with Dynamic motion primitives (DMP). The DMP approach allows simple adaptation of the demonstrated trajectory to the robot dynamics. In the second approach, we use reinforcement learning, which allows learning without any previous knowledge of the system or the environment. In contrast to the majority of the previous attempts, we used SASRA learning algorithm. Experimental results for both cases were performed on Mitsubishi PA10 robot arm.

Recently (2011) we applied also error learning in constrained domain. The search domain was constrained with previous human demonstrated trials (successful or unsuccessful). Using this approach, the robot learned the appropriate policy in only 4 trials.

Multimedia

Learning of Ball-in-a-cup game exploiting previous experience (KendamaLearning.avi, 3923 kb)
Reinforcement learning (Reinforcement.avi, 3485 kb)
Imitation learning (Imitation.avi, 1685 kb)
Simulation of reiforcement learning (SwingUp250T.avi, 6296 kb)

Publications

Inproceedings

  • Nemec B., Vuga R., Ude A., Exploiting previous Experience co Constrain robot sensorimotor learning, 2011. [More] [PDF]
  • Nemec B., Zorko M., Žlajpah L., Learning of a ball-in-a-cup playing robot, 2010. [More] [PDF]