Quick Facts

Abstract

Over the last few decades, there has been an impressive amount of published work on legged locomotion, including bipedal walking, running, hopping, stand-ups, summersaults and much more. However, despite all this progress, legged locomotion research has largely been driven by researchers using human insight and creativity in order to generate locomotion control algorithms. In order to improve the robustness, energy efficiency, and natural appearance of legged locomotion, there may be a significant advantage to using machine learning methods to synthesize new controllers and to avoid tedious parameter tuning. For instance, it could be advantageous to learn dynamics models, kinematic models, impact models, for model-based control techniques. Imitation learning could be employed for the teaching of gaits patterns, and reinforcement learning could help tuning parameters of the control policies in order to improve the performance with respect to given cost functions. In this context, we would like to bring together researchers from both the legged locomotion and machine learning in order to discuss which locomotion problems require learning, and to identify the machine learning methods that can be used to solve them.

Goal

In order to better understand the application of machine learning techniques to locomotion, our goal is to bring together researchers who represent many different approaches to biped locomotion control with their peers in machine learning for control. We hope to discuss future research directions for principled machine learning approaches to biped locomotion. The workshop will address topics such as:

• Which unsolved biped locomotion problems can be solved using learning?
• Can walking be broken down into components upon which machine learning methods are applicable?
• What models (e.g., forward, inverse, impact) would be desirable for controlling locomotion?
• Can machine learning methods help solve the gait generation and foot-placement problems?
• Can human learning of locomotion yield insights for both robotics and machine learning?
• Which machine learning algorithms are suitable for online implementation on the robot, and which problems can be solved in simulation?
• What cost functions should be used to describe "optimal" walking, and what experiments should be done to test our controllers?

Furthermore, we intend to kick-off the Legged Robot Control Competition.

Program

Participants

This workshop will bring together researchers from both the robotics and machine learning in order to explore how to approach the topic of learning legged locomotion in a principled way. Participants of the workshop (inclusive of the audience) are encouraged to actively participate by responding with questions and comments about the talks and give stand-up talks. Please contact the organizers if you would like to reserve apriori some time for expressing your view (short short-talk!!) on a particular topic.

Organizers

The workshop is organized by Jan Peters, Russ Tedrake and Stefan Schaal, from the Departments of Computer Science and Neuroscience, University of Southern California, Los Angeles, CA, USA and from the Brain and Cognitive Sciences Department at the Massachusetts Institute of Technology, Cambridge, MA, USA.

Location and More Information

The most up-to-date information about Robotics - Science and Systems 2005 can be found on the Robotics 2005 website.