Integrating Learning from Demonstrations with Neuroadaptive Control for Robotic Trajectory Tracking

Learning from Demonstrations (LfD) has the capability to transfer human expertise on repetitive tasks into robotic systems, thus avoiding the need for highly technical knowledge to program such systems. A common formulation for modeling human demonstrations as non-linear trajectories, as well as for ensuring robustness against perturbations, is state-dependent dynamical systems (DSs). Recent DS-based LfD approaches learned the complex dynamics of motion and provided stability guaranties; however, most of them followed the next integral curves of the DS to reach their target. These approaches are unsuitable for applications that require precise tracking of the robot's trajectory. To address this drawback, this study proposes a neuroadaptive control approach to enhance the tracking fidelity of learned trajectories, which provides performance guarantees online in a DS-based LfD approach. Furthermore, a constrained optimization problem based on the Gaussian Mixture Model (GMM) and the Control Lyapunov Function (CLF) is used to generate the reference trajectory offline. The proposed approach has been experimentally validated on the LASA dataset and on real trajectories coming from the Unmanned Surface Vessel (USV) Vendaval. Preliminary results confirm that the novel DS-based LfD approach proposed in this study significantly improves trajectory tracking when the system is disturbed; moreover, these results outperform existing approaches in terms of tracking fidelity.