Real-time neural network control of a biped walking robot

Cerebellar model arithmetic computer (CMAC) neural networks can be applied to the problem of biped walking with dynamic balance. The project goal here is to develop biped control strategies based on a hierarchy of simple gait oscillators, PID controllers, and neural network learning, that do not require detailed kinematic or dynamic models. While results of simulation studies using two-dimensional biped simulators have appeared previously, the focus in this article is on real-time control studies using a ten axis biped robot with foot force sensing. This ongoing study has thus far produced several preliminary results toward efficient walking. The experimental biped has learned the closed chain kinematics necessary to shift body weight from side-to-side while maintaining good foot contact, has learned the quasi-static balance required to avoid falling forward or backward while shifting body weight from side-to-side at different speeds, and has learned the dynamic balance required in order to lift a foot off of the floor for a desired length of time, during which the foot can be moved to a new location relative to the body. Using these skills, the biped is able to march in place and take short steps without falling (too often).<>