A kinodynamic planning-learning algorithm for complex robot motor control

Abstract

Abstract-Robot motor control learning is currently one of the most active research areas in robotics. Many learning techniques have been developed for relatively simple problems. However, very few of them have direct applicability in complex robotic systems without assuming prior knowledge about the task, mainly due to three facts. Firstly, they scale badly to continues and high dimensional problems. Secondly, they need too many real robot-environment interactions. Finally, they are not capable of adapting to environment or robot dynamic changes. In order to overcome these problems, we have developed a new algorithm capable of finding from scratch open-loop state-action trajectory solutions by mixing sample-based tree kinodynamic planning with dynamic model learning. Some results demonstrating the viability of this new type of approach in the cart-pole swing-up task problem are presented.

RoMPLA: An efficient robot motion and planning learning architecture

“…Nonetheless, the scalability of this class of algorithms to high dimensional state and action spaces is still a matter of study, more in the case of continues state-action spaces. To overcome this problem, the authors of this publication proposed in previous work a new model-based learning algorithm, called KiPLA [8], which handles much more efficiently the curse of dimensionality problem. This algorithm mixes kinodynamic planning and model learning for the purpose of finding suboptimal open-loop policies for achieving a certain task.…”