Low-Dimensional Learning for Complex Robots

This paper presents an algorithm for learning the switching policy and the boundaries conditions between primitive controllers that maximize the translational movements of a complex locomoting system. The algorithm learns an optimal action for each boundary condition instead of one for each discretized state-action pair of the system, as is typically done in machine learning. The system is modeled as a hybrid system because it contains both discrete and continuous dynamics. With this hybridification of the system and with this abstraction of learning boundary-action pairs, the “curse of dimensionality” is mitigated. The effectiveness of this learning algorithm is demonstrated on both a simulated system and on a physical robotic system. In both cases, the algorithm is able to learn the hybrid control strategy that maximizes the forward translational movement of the system without the need for human involvement.