How to Compensate Friction in Dynamic Hard Contact Tasks by Means of Neural Networks

We present a new method of neural friction compensa tion in manipulator control especially during dynamic tasks with de ned contact to external environment Such manipulator movements involve internal joint friction and external friction between tool and environment The suggested method compensates friction caused dis turbances by means of neural networks Based on a minimal friction model Lyapunov stability theory is used to gain a stable learning rule for Radial Basis Function RBF neural networks This novel compen sation method is of particular importance in force position control for slow movements in which static friction and stick slip friction are the foremost e ects of disturbance We implemented this method in a real time control of an DOF industrial manipulator Siemens manutec r Results from experiments are presented Using additional RBF networks to adapt to di erent kinds of e ects e g mass coupling gravitation load mass this algorithm can easily be extended to unknown manipulators Introduction In actual applications robot manipulators are expected to perform sophisticated tasks under con strained conditions To realize tasks like controlling the end e ectors orientation and movement in contact with unknown hard surfaces the control system has to be exible enough to cope with various disturbances e g joint friction tool friction uncertain model parameters We developed an adaptive force control for industrial tasks like deburring chamfering or robotic based thermoplastic bre placement where normal desired forces have to be exerted on unknown shaped surfaces and friction e ects due to slow movements or high contact forces are urgent problems Coulomb friction is uncertain and can vary signi cantly with load wear and temperature To get better performance in force and trajectory tracking with the developed force position control a compensation method for adaptive friction compensation satisfying these requirements is presented Theory In this work Lyapunov stability theory was used to gain a stable learning rule for Radial Basis Function RBF neural networks from the derivative V x where V x is the Lyapunov energy function of the system Considering the manipulator dynamics for free movements M C   g Fv FC sign as a conservative system with dissipative terms friction and a control law composed from di erent kinds of analytical and neural nonlinear compensations adaptation rules for analytic parameters and learning rules for the RBF networks were derived Our compensation method is an extension of similar This work was supported in part by Federal Ministry for Education Science Research and Tech nology BMBF under grant DEMON