Compliant robot control using cerebellar spiking neural networks, a biologically inspired approach

Abstract

In the last decades, a new robotics paradigm has been introduced due to physical human-robot interaction (HRI) and the use of collaborative robots (cobots) equipped with low-power actuators and elastic components. This scenario requires the use of cobot controllers able to operate in unstructured environments and that do not depend on the accurate mathematical modeling of the nonlinear dynamics introduced by elastic elements. Robot behavior in this context is required to emulate the adaptability and flexibility of human behavior as much as possible. The cerebellum, pivotal for human motor control, has long been proposed as an adaptive controller, and its regular neural structure has allowed the development of computational models which replicate, to some extent, its structural and functional properties. Here, we propose a cerebellar-based adaptive controller able to provide torque control of a cobot with nonlinear dynamics. Using spiking neural networks we replicate the cerebellum neural topology and synaptic plasticity mechanisms. We then embed the biologically plausible cerebellar network at the core of a cobot control loop. The spike-processing computational cost of biologically plausible cerebellar models has prevented their real-world applicability, thus relegating them to mere theoretical or simulated models. Within this dissertation, we prove the applicability of our biologically plausible cerebellar controller in real-world control problems. We present a cerebellar spiking neural network which is large enough to provide the required resolution for torque control of six degrees of freedom in real-time, and hence can operate real cobots. The cerebellar controller provides fine accuracy in the execution of different motor tasks thanks to the deployed cerebellar learning mechanisms. Besides, the controller is also able to adapt the cobot behavior to unstructured changes directly affecting the cobot dynamics. Furthermore, the aforementioned cerebellar control learning mechanisms can also cope with sensorimotor delays affecting the robotcontroller communication, a well-known source of control loop instability. Sensorimotor latency is unavoidable in the central nervous system (CNS), however, it does not jeopardize the stability of motor control thanks to, among others, cerebellar predictive behavior. We prove the cerebellar controller robust against sensorimotor delays of different nature, thus applying to robotics another intrinsic feature of the cerebellum. In addition to cerebellar control, we expand the biologically inspired approach with other key elements of the CNS and musculoskeletal system. We present some first results of adding spinal cord circuits to the cerebellar controller. The spinal cord, using direct muscle feedback to allow fast-stretch reflexes and muscle activity regulation, is found to improve cerebellar learning and robustness against perturbations. As next step we will integrate the spinal cord circuits and the cerebellar controller operating the cobot, for which muscle dynamics will need to be added to the control loop. Here we present a preliminary approach for the integration of muscle dynamics within the cobot control loop, which is shown capable of modifying the motion stiffness of the cobot by changing the cocontraction degree of antagonistic muscle pairs. Different stiffness profiles would allow the robot behavior to cover different degrees of admittance and impedance control, of interest to physical HRI as those control modes directly impact how the robot reacts to external interactions (admittance control performs better in soft environments, while impedance control favors stiff environments). For collaborative robotics to succeed, robot performance must emulate the adaptability and flexibility of human behavior. Hence, the biological substrate behind human conduct could be used as inspiration to bring robot behavior closer to our inherent motor capabilities. Human behavior is sustained by both hardware and software: the biomechanics of the musculoskeletal system together with the control provided by the CNS allow us to interact with others and the environment. On the hardware side, robot design is increasingly mimicking the dynamics of living beings. On the software side, the study and understanding of the different CNS areas and their computational replication can expand the family of controllers able to provide adaptive, compliant robot control. Here, we benefit from decades of neuroscience studies about the cerebellum structure and functioning, and apply those findings to current robotic challenges.