Bio-inspired motor learning models for robot control

Resumen

We propose two bio-inspired architectures for representing the role of the adaptive cerebellar microcircuit in correcting the motor behavior based on current errors: the feedforward and the recurrent schemes. While in the first architecture, the cerebellum adds adaptive torque corrections and its weights are adjusted depending on the motor-error signal, in the second one it adds sensory correction terms to the controller and its weights depend on the sensory error. Biological systems perform adaptation of dynamics and kinematics models for accurate control with low power actuators, while in robotics, robots normally achieve very high precision and high speed motion with high forces and high energy consumption. This industrial approach cannot be used in the framework of human interaction applications because it is potentially dangerous in case of malfunctioning. Thus, adoption of compliant strategies of dynamics and kinematics models abstraction and adaptive control schemes in real time in the field of robotics are required. Furthermore, the cerebellum is able to acquire intrinsic models through experience by a perceptual feedback that allows the motor learning to proceed. These models are called internal models. Motivated by this, we implemented a cerebellum model able to adapt its corrections and store the sensory consequences or feedforward motor commands for predicting appropriate actions when needed. We addressed this study with a machine learning approach (LWPR algorithm) embedded in the control loops in which the LWPR module abstracts the functionality of the input layer to the cerebellar cortex. LWPR provides optimal input representation to the granular layer in terms of neural resources, that is, it adapts its neural resources incrementally and according to the input data structure. Moreover, we built compliant control systems combining the feedback error learning approach and adaptive predictive control. This dissertation first introduces a description of the cerebellar microcircuitry and of the role of internal models in motor control. Then, we present the control systems based on inverse and forward internal model for controlling a robot arm in which a feedback error controller leads to a precise, compliant and stable control during manipulation of objects and the cerebellar-machine learning synergy makes the robot adaptable to changing conditions. Finally, we show the principal results obtained of the performance evaluation of both systems. We demonstrate the validity and efficiency of the models with experiments on a 3-DOF and 7-DOF light-weight robot arm.