Learning RoboCup Skills

From PRLT
Jump to: navigation, search
MRTLogo.jpg
RoboCupLogo.jpg


This project aims at applying Reinforcement Learning techniques for learning both low- and high-level skills for RoboCup robots.

Introduction

Ongoing work

Go To Ball Behavior

Triscar.jpg

The Go To Ball is one of the simplest behaviors for a soccer robotic player. Although simple to implement by hand, it introduces many difficulties for the traditional RL algorithms. In particular, the continuous state and action space cannot be dealt with only through simple discretization because the learning process would be either too long or too unreliable (some details about the issues in learning in robotics applications can be found here). In this first experiment we tried to overcome these problems by using a novel RL algorithm: PWC-Q-learning (Piecewise Constant Q-learning).

Resources

GoToBall Movies:

High Resolution Movie

Low Resolution Movie

MRT Official Website

People involved

  • Alessandro Lazaric, PhD Student
  • Marcello Restelli, PhD
  • Moreno De Amicis, Master Student

Publications

A. Bonarini, A. Lazaric, M. Restelli - Piecewise Constant Reinforcement Learning for Robotic Applications
4th International Conference on Informatics in Control, Automation and Robotics (ICINCO) , 2007
Bibtex
Auteur : A. Bonarini, A. Lazaric, M. Restelli
Titre : Piecewise Constant Reinforcement Learning for Robotic Applications
Dans : 4th International Conference on Informatics in Control, Automation and Robotics (ICINCO) -
Adresse :
Date : 2007