Abstract: Robot Learning from Demonstration (RLfD) is a paradigm that seeks to enable users to teach personal robots arbitrary tasks, allowing robots to better serve users' wants and needs without explicit programming. We apply RLfD to the task of robot soccer, a behavior commonly manually coded as part of the RoboCup games. Specifically, we attempt to learn the 'swarm-style' goal scorer shown above. This task, framed as a Finite State Machine, causes issues in regression-based learning without explicit transition information, as the underlying mapping from perception to actuation is multimodal. We are developing nonparametric Bayesian techniques for learning multimaps in an incremental, sparse fashion, suitable for interactive robot tutelage.

---

Related Publications:

• Learning multi-objective control policies from demonstration. Daniel H Grollman and Odest Chadwicke Jenkins.
  In IROS workshop on Robotics Challenges for Machine Learning, Nice, France, September 2008.
  [ bib | Poster | Text ]
• Incremental nonparametric bayesian regression. Frank Wood, Dan H. Grollman, K. A. Heller, Odest Chad Jenkins, and Michael Black.
  Technical Report CS-08-07, Brown University Department of Computer Science, 2008.
  [ bib | Text ]
• Sparse incremental learning for interactive robot control policy estimation. Daniel H Grollman and Odest Chadwicke Jenkins.
  In International Conference on Robotics and Automation, pages 3315-3320, Pasadena, CA, USA, May 2008.
  [ bib | Slides | Poster | Text ]
• (Machine) learning robot control policies. Daniel H Grollman and Odest Chadwicke Jenkins.
  In NIPS Workshop on Robotics Challenges for Machine Learning, Whistler, BC, Canada, December 2007.
  [ bib | Video | Poster | Text ]
• Learning robot soccer from demonstration: Ball grasping. Daniel H Grollman and Odest Chadwicke Jenkins.
  In R:SS Workshop on Robot Manipulation: Sensing and Adapting to the Real World, Atlanta, GA, USA, June 2007.
  [ bib | Poster | Text ]
• Learning robot soccer skills from demonstration. Daniel H Grollman and Odest Chadwicke Jenkins.
  In International Conference on Development and Learning, pages 276-281, London, UK, July 2007.
  [ bib | Teaser | Poster | Text ]
• Dogged learning for robots. Daniel H Grollman and Odest Chadwicke Jenkins.
  In International Conference on Robotics and Automation, pages 2483 - 2488, Rome, Italy, April 2007.
  [ bib | Video | Slides | Text ]

---

Videos:

• Game-based learning. Daniel Byers, Michael Lapping-Carr, Julie Kumar, Thea Hinkle, Dan Grollman, and Chad Jenkins.
  AAAI2008 Video Competition, July 2008. Winner: Best Student Video.
  [ bib | Video ]
• Multimap goal scoring (GS)
  [Demonstrated | Learned | oldExample]
• Unimap goal scoring (GSF)
  [Demonstrated | Learned]
• Learning Goal-Charging from Human Demonstration
  [MP4]
• Overview of the learning system
  [MP4 | ICRA2007]
• The Ball AcQuire Task (AQ)
  [Example | Demonstrated | Learned]
• The Ball Track Task (BT)
  [Example | Demonstrated | Learned with LWPR | Learned with SOGP | Learned from Human Data with SOGP]
• The Head-Tail Task (HT)
  [Example | oldExample]
• The TRap Task (TR)
  [Example]
• The Ball-Approach Task (BA)
  [Example]
• The Walk Task (WK)
  [Example]
• The RPlay project.
  [vid]

---