A Modular Bio-Inspired Architecture for Motor Learning and Control of Robotics Systems

Classical robotic control methods struggle in overcoming the constraints and challenges of modern robotics applications. Nowadays, robots require a high level of flexibility to adaptively work in a wide range of scenarios. Our study proposes robotic control solutions that take inspiration from the vertebrates’ central nervous system (CNS) to endow robots with the necessary adaptive and predictive capabilities. In the thesis, we first shed light on the neural mechanisms employed by the CNS to produce complex motor movements in dynamically challenging conditions. Among all the CNS regions involved in motor control, the research focuses on the cerebellum, a powerful and compact neural circuit well known for its crucial role in adaptive learning and control of complex motor behaviors. Based on the challenges and considerations identified from the review of the literature, we propose distinct biologically inspired architectures for robotic real-time adaptive motor learning and control in unknown and disturbed environments. The cerebellar-like control schemes embed a cerebellar-like simulations model that aims to artificially reproduce the functionality, plastic learning, modularity, and morphology of the cerebellum through the combination of machine learning, artificial neural networks, and computational neuroscience techniques. The cerebellar-like control schemes mimic through engineering techniques the different theories regarding the acquisition and employment of cerebellar internal models for the control of robotic motor behavior in dynamically changing conditions. The research merges ideas proposed by the scientific community in the last decades into a unique system that is suitable for real-time robotic applications and attempts to answer through robotics experiments various scientific assumptions regarding the cerebellar internal models theories. The empirical results show the incredible contribution that a cerebellar-like system can incorporate whether the robotic control architecture is affected by high modeling errors, unobservable and high dimensional state and action spaces, uncertainties, sensor noise, external perturbations, and changes in the dynamics. Even though there are many ongoing discussions regarding how the cerebellum operates, we believe that the extraordinary potential of the cerebellar-like methods can endow robots with the flexibility and dynamism that modern robotic applications require.