We describe our approach to the robot's Hanetsuki task (Japanese badmi
nton), i.e. to return the incoming ball to the human opponent with a r
acket. A learning algorithm that consists of updating action commands
and smoothing them based on a Gaussian kernel is proposed to compensat
e for the insufficiency in a non-adaptive model-based approach. Experi
mental results obtained by using the developed Hanetsuki robot are als
o shown.