Direct Neural-Adaptive Control of Robotic Manipulators using a Forward Dynamics Approach

This paper uses a forward-dynamics approach to achieve direct neural-adaptive control of a two-link robotic manipulator. Cerebellar model articulation controllers model the forward dynamics. Previous approaches in the literature use an inverse-dynamics approach because online estimation of the inertia matrix is difficult. The proposed method succeeds by using a supervisory inertia matrix when updating the neural network weights. The supervisory matrix does not need to accurately model the real inertia matrix to achieve accurate trajectory tracking. This remains true even when significant unmodelled payloads are added or, equivalently, when there is large uncertainty in the inertia matrix. A Lyapunov analysis establishes the ultimate uniform boundedness of all signals