A hierarchical primitive-based learning tracking framework for unknown observable systems based on a new state representation

A three-level learning-based hierarchical tracking framework is validated in this work and aims at endowing control systems with generalization capabilities specific to artificial intelligence. The framework operates at three levels: the low-level L1 is concerned with ensuring a linear time-invariant (LTI) behavior from the reference input to the controlled output in terms of reference tracking. The second level L2 acts on top of the linearized closed-loop dynamics, to learn primitive pairs (reference inputs-controlled outputs pairs), in an entirely experimentally-driven style, using Iterative Learning Control (ILC). These primitives are optimally learned for tracking a desired trajectory and they naturally learn by trials, due to the ILC principle. Finally, the third level L3 uses the learned primitive pairs to extrapolate the optimal behavior to new desired trajectories, without relearning by trials. Level L3 is able to predict the optimal reference inputs that ensure accurate tracking in new tracking scenarios, which is a feature specific to intelligent beings. The framework is validated on a multivariable nonlinear two-joints rigid planar manipulator.