Learning to Walk through Imitation

CREATED: 200711070834 Speaker: Rawichote Chalodhorn WWW: http://neural.cs.washington.edu/people/misc/choppy/Site/Research.html

** Learning to Walk Through Imitation

• motivation: humanoid robots in RoboCup
• hand tuned motions is timed consuming, depends on specific robots
• learning from motion capture ** differences between human and robot model ** needs to be optimized to match dynamics of robot ** direct optimization of high dimensional data is intractable ** dynamic model of robot not available

** Approach

• reduce the number of dimensions using dimension reduction techniques
• further reduce the amount of data by modeling the points in the posture subspace as a loop in cylindrical coordinate system
• learn a model of the gyroscope signals which indicate stability of the robot, using a time delay RBF network
• posture command + history of gyroscope signals -> new gyroscope signals at time t
• optimize the posture to minimize the gyroscope signal (instability)
• may need additional iterations to provide feedback to improve the signal predictor model
• motion scaling: scale down motion capture information to reduce the amount of movement, learn the motions from scaled down motion capture data first and scale up gradually
• does not need dynamic model, can be applied to other robots