(:quicktime http://www-clmc.usc.edu/movies/imitation/DA-PoleBalancing-Imitation.mov scale=aspect width=120 height=110 autostart="false" :)

The Sarcos Dexterous Arm learns the skill of pole balancing and a pendulum-swing-up from a demonstration of a teacher. All appled learning methods involve model-based reinforcement learning.

Using a novel methods of encoding movement primitives as nonlinear dynamical systems, the humanoid robot DB is taught how to perform tennis forehand and backhand swings. Note that in imitation, the movement is performed towards the colored ball, i.e., the movement primitive is generalized to new targets.

(:quicktime http://www-clmc.usc.edu/movies/imitation/Imitation-Discrete.mov scale=aspect width=120 height=110 autostart="false" :)

(:quicktime http://www-clmc.usc.edu/movies/imitation/Imitation-Rhythmic.mov scale=aspect width=120 height=110 autostart="false" :)

As in the previous movie, dynamical systems are used to encode movement primitives, just that we show how to encode rhythmic behavior this time. Note that the learned primitives can be sped up easily, can be changed in amplitude, and can be made robust towards external perturbations. The latter point is demonstrated by manually interfering with the robot's movement, which leads to a temporary abortion of the movement pattern. After the robot is released, it continues the pattern where it was stopped -- i.e., the robot planning was gracefully interupted.