A short time ago, cognitive science research and autonomous robotics were
utterly unrelated subjects, requiring completely different background, skills
and methodologies. Nowadays, the distance between the two fields is being
constantly shortened by the progress in computational modeling, and the construction
of increasingly skilled autonomous artificial agents inspired by the
abilities and behavior of living beings. The astounding discoveries that are
being recently achieved by brain scientists constitute the fundamental building
blocks for computational neuroscience and biomimetic robotics. This thesis
presents interdisciplinary research which puts neuroscience and robotics on the
same level, aiming at their mutual enrichment.
Grasping and manipulation of every kind of object is arguably the most distinctive
practical skill of human beings, and erect posture has likely evolved in
order to free the upper limbs and make of the hands two unmatchable tools.
Despite the great efforts that have been and are being put on it, grasping in
robotics is largely an unsolved problem, due to its inherent complexity and the
still limited adaptive skills of present day robots in visual and visuomotor behaviors.
The task of object grasping is dealt with in this thesis by mimicking,
as accurately as possible, the brain mechanisms which underlie planning and
execution of grasping actions in humans and other skilled primates.
The principal contribution of this thesis is the definition and implementation
of a functional model of the brain areas involved in vision-based grasping actions.
The model constitutes a bridge between cognitive science and robotics
research, and includes all the steps required for performing a successful grasping
action from visual data. The subdivision of visual processing into the
dorsal and ventral cortical streams, respectively dedicated to action-oriented
and perception-oriented vision, is thoroughly taken into account. Hypotheses
regarding the mechanisms that allow to achieve complex interactions with
the peripersonal space, through the integration of the data provided by the
streams, are put forth. Transfer functions are proposed for modeling the visuomotor
transformations performed by the brain areas most critical in grasp
planning and execution. Object shape and pose estimation takes into account
and integrates the contributions of stereoptic and perspective visual cues.
The particular attention payed to the functional role of brain areas makes the
model especially suitable for implementation on a real robotic setup, and a full
vision-based robotic grasping system has been developed following its guidelines.
Visual information regarding an unknown target object is acquired and
transformed into a basic representation onto which two concurrent processing
mechanisms are performed. The dorsal stream extracts and analyzes possible
grasping features, while the ventral stream performs object classification.
Dorsal and ventral visual data are merged for estimating shape, size and position
of the object, and a grasping plan is devised which takes into account
both visual data and proprioceptive information on the state of the arm and
hand. Grasp execution is performed with the aid of tactile feedback in order
to achieve a stable final hand configuration.
Grasping experiments have been performed on real objects unknown to the
system, and the obtained results attest the achievement of the thesis’ two
concurrent goals. On the one hand, the system can safely perform grasping
actions on different unmodeled objects, denoting especially reliable visual and
visuomotor skills. This confirms that the new research path proposed by the
thesis, according to which robotic grasping can be based on the integration
of the two visual processing channels of the primate brain, is significant and
worth further exploration. On the other hand, the computational model and
the robotic experiments help in validate theories on the mechanisms employed
by the brain areas more directly involved in grasping actions. This thesis offers
new insights and research hypotheses regarding such mechanisms, especially
for what concerns the interaction between the streams. Moreover, it helps in
establishing a common research framework for neuroscientists and roboticists
regarding research on brain functions.
A short time ago, cognitive science research and autonomous robotics wereutterly unrelated subjects, requiring completely different background, skillsand methodologies. Nowadays, the distance between the two fields is beingconstantly shortened by the progress in computational modeling, and the constructionof increasingly skilled autonomous artificial agents inspired by theabilities and behavior of living beings. The astounding discoveries that arebeing recently achieved by brain scientists constitute the fundamental buildingblocks for computational neuroscience and biomimetic robotics. This thesispresents interdisciplinary research which puts neuroscience and robotics on thesame level, aiming at their mutual enrichment.Grasping and manipulation of every kind of object is arguably the most distinctivepractical skill of human beings, and erect posture has likely evolved inorder to free the upper limbs and make of the hands two unmatchable tools.Despite the great efforts that have been and are being put on it, grasping inrobotics is largely an unsolved problem, due to its inherent complexity and thestill limited adaptive skills of present day robots in visual and visuomotor behaviors.The task of object grasping is dealt with in this thesis by mimicking,as accurately as possible, the brain mechanisms which underlie planning andexecution of grasping actions in humans and other skilled primates.The principal contribution of this thesis is the definition and implementationof a functional model of the brain areas involved in vision-based grasping actions.The model constitutes a bridge between cognitive science and roboticsresearch, and includes all the steps required for performing a successful graspingaction from visual data. The subdivision of visual processing into thedorsal and ventral cortical streams, respectively dedicated to action-orientedand perception-oriented vision, is thoroughly taken into account. Hypothesesregarding the mechanisms that allow to achieve complex interactions withthe peripersonal space, through the integration of the data provided by thestreams, are put forth. Transfer functions are proposed for modeling the visuomotortransformations performed by the brain areas most critical in graspplanning and execution. Object shape and pose estimation takes into accountand integrates the contributions of stereoptic and perspective visual cues.The particular attention payed to the functional role of brain areas makes themodel especially suitable for implementation on a real robotic setup, and a fullvision-based robotic grasping system has been developed following its guidelines.Visual information regarding an unknown target object is acquired andtransformed into a basic representation onto which two concurrent processingmechanisms are performed. The dorsal stream extracts and analyzes possiblegrasping features, while the ventral stream performs object classification.Dorsal and ventral visual data are merged for estimating shape, size and positionof the object, and a grasping plan is devised which takes into accountboth visual data and proprioceptive information on the state of the arm andhand. Grasp execution is performed with the aid of tactile feedback in orderto achieve a stable final hand configuration.Grasping experiments have been performed on real objects unknown to thesystem, and the obtained results attest the achievement of the thesis’ twoconcurrent goals. On the one hand, the system can safely perform graspingactions on different unmodeled objects, denoting especially reliable visual andvisuomotor skills. This confirms that the new research path proposed by thethesis, according to which robotic grasping can be based on the integrationof the two visual processing channels of the primate brain, is significant andworth further exploration. On the other hand, the computational model andthe robotic experiments help in validate theories on the mechanisms employedby the brain areas more directly involved in grasping actions. This thesis offersnew insights and research hypotheses regarding such mechanisms, especiallyfor what concerns the interaction between the streams. Moreover, it helps inestablishing a common research framework for neuroscientists and roboticistsregarding research on brain functions.
การแปล กรุณารอสักครู่..
