Organizers:

• Jane Mulligan, University of Colorado, Boulder, CO, USA.
• Greg Grudic, University of Colorado, Boulder, CO, USA.

Workshop Description:

Robots have been successful in tasks where they can be rigidly programmed for highly structured environments like factory assembly lines. However, the dream of robots that work alongside or in lieu of people in natural environments has long evaded researchers in Artificial Intelligence. As an example, robots cannot autonomously navigate through many types of outdoor environments (as supporting evidence, see the DARPA Grand Challenge - http://www.darpa.mil/grandchallenge ). Although, progress has been made on robots that move in stable man-made environments such as offices (see http://www.evolution.com for an example), many applications such as search and rescue, security or even elder care could be addressed if robots could operate robustly in unknown, dynamic and unstructured spaces.

In contrast, human operators can often remotely control a robot to accomplish complex tasks in a variety of environments, using only the robot's sensors and actuators (see http://www.omnitech.com/pdf/sts_ds.pdf ). This implies that the robot's sensors and actuators are not to blame, it is the controllers we write that are inadequate.

The goal of this one day workshop is to investigate and propose Machine Learning based approaches to autonomous robotic problems in unstructured environments. One example of such challenging domains is exemplified by the DARPA LAGR program, where the goal is robust navigation in outdoor environments. This task is characterized by a very high dimensional input space which includes four high resolution color cameras, IR sensors, inertial sensors, odometry, and GPS. Recent developments in Machine Learning for complex domains may give insight into Robotics in general. Specifically, this workshop will address, among others, the following topics:

• The use of human teleoperation data to learn autonomous robot controllers.
• Manifold mapping techniques for identifying the low dimensional sensor and actuator space where the robot operates. Examples include unsupervised learning algorithms such as LLE, ISOMAP, and Spectral Clustering.
• Semi-supervised learning techniques for reducing the number of training examples required to learn autonomous robot controller models.
• Reinforcement Learning algorithms for very high dimensional spaces where the number of rewards received is very limited.
• Identification of what features (extracted from raw sensory data) are useful for tasks in unstructured environments.

Call for Submissions:

High quality submissions on the topics above or related are encouraged. We tentatively plan to follow up the workshop with a special journal issue or a book.

Submissions should be in JMLR paper format (see http://jmlr.csail.mit.edu/format/format.html ), and should be no longer than 10 pages. Submissions should be in PDF format and emailed by Oct 19, 2005 to .

Important Dates:

• Oct. 19, 2005 : Paper Submission (midnight MDT) EXTENDED!
• Nov. 15, 2005 : Author notification
• Dec. 10, 2005 : Workshop at Whistler

• Morning session: 7:30am--10:30am

• 7:30am Opening Remarks
  Greg Grudic and Jane Mulligan
• 7:40am DARPA LAGR
  Larry Jackel
• 8:10am Representing Natural Objects in Unstructured Environments
  Fabio Ramos, Suresh Kumar, Ben Upcroft, Hugh Durrant-Whyte
• 8:40am Learning for Autonomous Navigation: Extrapolating from Underfoot to the Far Field
  L. Matthies, M. Turmon, A. Howard, A. Angelova, B. Tang, E. Mjolsness
• 9:10am coffee break
• 9:25am Discovering Natural Kinds of Robot Sensory Experiences in Unstructured Environments
  Daniel H. Grollman, Odest Chadwicke Jenkins, Frank Wood
• 9:40am Learning Landmarks for Localization via Manifolds
  Xuming He, Volodymyr Mnih, Richard S. Zemel
• 9:55am Spotlight -- 5 min poster intros
  Relating Reinforcement Learning Performance to Classification Performance
  John Langford and Bianca Zadrozny
  • Probabilistic Robot Planning Under Model Uncertainty: An Active Learning Approach
    Robin Jaulmes, Joelle Pineau, Doina Precup}
  • Region-Based Value Iteration and its Application to Robot Navigation in a Minefield
    Hui Li, Lihan He, Xuejun Liao, Shihao Ji, Lawrence Carin
  • Learning Qualitative Markov Decision Processes
    Alberto Reyes, L. Enrique Sucar, Eduardo F. Morales, Pablo H. Ibarguengoytia
  • Behavioural Cloning for Robots in Unstructured Environments
    M. Waleed Kadous, Claude Sammut, Raymond Ka-Man Sheh
  • Efficient Sample Reuse by Off-policy Natural Actor-critic Learning
    Takeshi Mori, Yutaka Nakamura, Shin Ishii
  • Socially Guided Machine Learning: Designing an Algorithm to Learn from Real-time Human Interaction
    Andrea Lockerd Thomaz, Cynthia Breazeal
  • Manifolds From Sensory Experience
    Greg Grudic, Jane Mulligan
  
  

• Afternoon session: 3:30--6:30pm

• 3:30pm Learned Optical Flow for Mobile Robot Control
  John Rogers, Andrew Lookingbill, Sebastian Thrun
• 4:00pm Learning Obstacle Avoidance Parameters from Operator Behavior
  Bradley Hamner, Sebastian Scherer, Sanjiv Singh
• 4:30pm Practical Techniques for Inference, Planning, and Learning in Autonomous Outdoor Navigation
  Paul Vernaza and Daniel Lee
• 4:45pm Maximum Margin Planning
  J. Andrew Bagnell, Nathan D. Ratliff, Martin A. Zinkevich
• 5:00pm coffee break
• 5:00pm Poster Session
• 5:30pm Panel Discussion