A hierarchical approach for adaptive humanoid robot control

In this paper, we propose a hierarchical approach called "CBR augmented GP" to evolve robust control programs for humanoid robots. Humanoid robots are high-dimensional systems; thus it is very difficult for GP to generate control programs for humanoid robots. The key idea in our approach is to extract control rules with GP in simplified simulation and get a prototype of the control program then interpret and interpolate it with Case-Based Reasoning (CBR) in the real world environments. Accordingly, our proposed approach consists of two stages: the evolution stage and the adaptation stage. In the first stage, the prototype of the control program is evolved based on abstract primitive behaviors in a highly simplified simulation. In the second stage, the best control program is applied to a physical robot thereby adapting it to the real world environments by using CBR. Experimental results show that this approach can generate robust control programs that can easily overcome gaps between simplified simulation and real world. Furthermore, the robot can adapt to new environments which it never encountered in simulation.