Christensen, David Johan4; Larsen, Jørgen Christian5; Stoy, Kasper5
1 Department of Electrical Engineering, Technical University of Denmark2 Automation and Control, Department of Electrical Engineering, Technical University of Denmark3 Centre for Playware, Automation and Control, Department of Electrical Engineering, Technical University of Denmark4 Copenhagen Center for Health Technology, Center, Technical University of Denmark5 University of Southern Denmark
This paper presents experiments with a morphology-independent, life-long strategy for online learning of locomotion gaits. The experimental platform is a quadruped robot assembled from the LocoKit modular robotic construction kit. The learning strategy applies a stochastic optimization algorithm to optimize eight open parameters of a central pattern generator based gait implementation. We observe that the strategy converges in roughly ten minutes to gaits of similar or higher velocity than a manually designed gait and that the strategy readapts in the event of failed actuators. We also optimize offline the reachable space of a foot based on a reference design but finds that the reality gap hardens the successfully transference to the physical robot. To address this limitation, in future work we plan to study co-learning of morphological and control parameters directly on physical robots.