Adaptive stiffness control of passivity-based biped robot on compliant ground using double deep Q network

Passive dynamic walking exhibits human-like and energy-efficient gait. Biologically inspired compliance introduced to flexible passivity-based robot would be helpful to generate stable locomotion. However, designing adaptive controller for flexible biped on compliant ground still remains a challenge. This paper aims to design an adaptive and model-free stiffness controller for passivity-based flexible biped on compliant ground, where the hip stiffness is modulated by double deep Q network. One benefit of the double deep Q network is that adaptive stiffness control policy could be directly learned from inputs. At first, passive dynamic walking gait is utilized as a reference trajectory during double deep Q network training. Then the trained double deep Q network is used as adaptive stiffness controller for biped on compliant ground. Simulation results show that the passivity-based biped robot could walk in such walking cases as disturbed initial condition, level compliant ground, downslope slippery compliant surface, and varying compliance environments. The adaptive stiffness controller would be used to make the passivity-based biped robot adapt to the environmental changes.