DTA

Digital Theses Archive

 

Tesi etd-05042018-165104

Type of thesis
Perfezionamento
Author
VANNUCCI, LORENZO
URN
etd-05042018-165104
Title
Brain-inspired methods for adaptive and predictive control of humanoid robots
Scientific disciplinary sector
ING-INF/06
Course
INGEGNERIA - Biorobotics
Committee
relatore Prof.ssa LASCHI, CECILIA
Keywords
  • brain-inspired control
  • gaze stabilization
  • humanoid robotics
  • neurorobotics
  • smooth pursuit
  • spinal cord
Exam session start date
;
Availability
parziale
Abstract
This thesis advances the field of control of humanoid robots, focusing on brain-inspired approaches to the problem.<br>Human beings are able to negotiate complex environments thanks to their advanced bodies and brains. The human body is a complex machine that is able to perceive the world through different sensors modalities and to perform fine complex motions with a high number of strong, redundant, compliant actuators. However, humans would not be able to perform such feats if it were not for the orchestration performed by the nervous system, and especially by the brain. The human brain is the most advanced of the animal kingdom and is capable of displaying cognitive capabilities as well as motor control strategies. In particular, the main feature of the human brain is being capable to process the massive amount of sensory information received and to use it to perform both cognitive and motor tasks, often at the same time. Concerning, motion control, the brain is able not only to generate descending motor signal to the muscles that perform coordinated fine movements, but also to combine proprioceptive and exteroceptive sensory information to adapt movements depending on different conditions, in a predictive and anticipative fashion. Predicting changes in the environment and reacting accordingly is a must-have feature for robots that need to interact with an unstructured and dynamic environment that has been built by humans for themselves. And such is the case for humanoid robots, that are supposed to be employed either as service assistants and in search and rescue scenarios.<br>This works focus on control strategies that take inspiration from the human brain and that show some of these predictive and adaptive capabilities. This has been achieved by employing \textit{neuro-controllers}, controllers that embed algorithms mimicking the role of brain areas, specifically related to the task at hand. In particular, two technologies have been considered to implement neuro-controllers: machine learning approaches and Spiking Neural Networks (SNN). The formers have been widely used in different contexts, from artificial intelligence to data mining, and for many years in robotics to complement classic control approaches. In the context of this work, two different controllers have been developed that make use of machine learning methods: a predictive visual tracking architecture and a gaze stabilization controller provided with sensory-motor anticipation. Conversely, SNN have been mostly used in neuroscientific context to perform detailed simulations of the human brain. Albeit their use in robotics have been limited, there is a rising interest in this kind of neural simulations that can take the bioininspiration of robotic controllers to the next level. Due to research in neuro-controllers with SNN being in its infancy, the robotics community lacked effective tools to develop them. As part of this work, in the framework of the Human Brain Project, a contribution to the development of a comprehensive tool for the modelling and testing of neuro-controllers embedding SNN (the Neurorobotics Plaform) has been given. Moreover, this tool has been employed to show the effectiveness of the approach with simple examples that include a basic neuro-controller for visual tracking and controller relying on a retinal model for visual perception. This thesis also shows preliminary results toward the implementation of complex control architectures with SNN, focusing on developing low level translation mechanism for actuators and sensors based on simulation of the spinal cord circuitry.
Files