Virtual Motoneuron Learning Network for Trajectory Generation of Legged Robot

Walking animals in nature have shown complex behavior tracks and motion patterns when adapting to their environment. In order to imitate these natural characteristics of animals in the motion system of legged robots, a virtual motoneuron learning network (VMLN) is constructed to convert goal-directed behavioral instructions into neural signal strength, thus generating signal processing mechanisms and behavioral responses similar to animals. The network is composed of a central pattern generator (CPG), a virtual motoneuron module (VMM), a foot motor nerve strength modulation module (FMM) and an error- based fast online learning mechanism called the dual integral learner (DIL). VMM receives rhythmic signals from CPG and generates different motions. These motions are controlled by global and local shape parameters of muscles activation and target guidance points. Then, the signals for driving the joint motors of the robot are obtained through the inverse kinematics module. Finally, we integrate the DIL into the virtual motoneuron network to ensure that the entire generated trajectory can be followed with low tracking error, so as to achieve more complex and flexible movements. The experimental results show that: 1) VMM can flexibly select the desired curve or surface from a group of curves or surfaces by modifying the shape parameters appropriately; 2) DIL can track the whole trajectory automatically, quickly and accurately, and realize fast learning of complex trajectory. Therefore, our proposed VMLN can generate active adaption and stress behavior according to signal feedback, and improve the motion agility, coordination and diversity of robots.