FRINK Linked Data Fragments server

Matches in SemOpenAlex for { <https://semopenalex.org/work/W2808875021> ?p ?o ?g. }

• W2808875021 abstract "This dissertation explores multiple ways of adding prior knowledgeto neural networks used as controllers in robotics. It can largely besplit into two parts. The first part of the dissertation focuses onadding prior knowledge to the gait generation, to spend less time inthe optimization process to find efficient solutions. The second partof the dissertation focuses on the use of morphological computationas prior knowledge in the generation of stable gaits for legged robots.In the introduction, we discuss in depth what is meant by prior knowledgein this context. We show how the concept of prior evidenceemerges naturally from the creation of a probabilistic framework for‘degree of belief’. We then discuss how this prior evidence can be usedin robotics using a paradigm built on an alternative view on computation,called morphological computation. We argue how this approachmakes a natural match for controlling compliant robots.There are multiple ways to add prior knowledge to neural networks.As a first step, in chapter 2 we explored data augmentation as a way toteach neural networks how to be invariant to affine image transformations.Image augmentations are a known way to have a convolutionalneural network learn this invariance in natural images. We improve onthis idea by putting the affine transform as a differentiable layer into8the neural network, thereby allowing the neural network to encodethis invariance explicitly, rather than to have to encode this implicitlyin the values of its parameters. The network is then able to transformimages as a special type of layer, next to convolutional or dense layers.We show that explicitly encoding this prior knowledge of affineinvariance into the architecture outperforms the previous method ofusing image augmentations.Next, in chapter 3 we move our focus to robotics and develop threedifferent gaits for the quadrupedal compliant robot Oncilla: a sinebasedapproach, a biologically inspired half ellipse approach and aspline-based approach. After comparing these approaches, we findthat the method based on biological gaits is the most efficient of thethree, especially at higher speeds. After this, we move our attentionto approaches for turning. We showe the importance of scapulae forturning in quadrupedal robots. We also show that to be able to optimizethe gaits without relying on a model, a lot of prior knowledge isneeded to keep the time required for gait optimization low.Consequently, in chapter 4 we evaluate whether transfer learningknown gaits to gaits for new situations improves the optimization process.We analyzed this by starting the optimization process for varioussetups with gait parameters which had already been optimized for flatterrain. We find that it indeed works in most cases, and at least didnot hurt the optimization process. We uncover that in this case, thereduced amount of exploration of the parameter space required beforethe parameters converges to an optimal solution is the reason for awarm start helping the optimization process. The optimization algorithmcan, therefore, find good solutions faster, and fine-tune theparameters longer for a better end performance.After this, we move our focus to morphological computation. As a firstaspect in chapter 5, we study morphological sensing, and more specifically,whether we can use general purpose sensors available on a smalllegged robot to classify the underground it is walking on. Since thedynamics of the robot change with the underground it walks across,it should be possible to infer this underground from the sensors monitoringthe body of the robot. Since we do not require any specializedsensors for the detection of the underground, we can argue that weare using the body of the robot as a resource of computation for theclassification. We can indeed classify the underground successfully in9most cases, both with supervised and unsupervised algorithms. In asecond part of chapter 5, we delve into which properties of the modelsare important for the correct classification. We find indications in ourdata that both memory and non-linearities are important aspects ofthis classification process and that they reinforce each other, whichprovides a starting point for the research in the next chapter.Since gaits of legged robots are typically on the eigenfrequencies oftheir morphology, the morphology can probably be used as a resourcefor computation to generate the control signals. This is a conceptcalled morphological control. In chapter 6 we are indeed able to movepart of the control onto the morphology and show that there is atrade-off between the memory aspects and the non-linear dynamicsneeded for it to perform well. It seems that the main parameteris the number of uncorrelated signals the linear regression receives.The more signals with information, the better the performance andthe smaller the error between the found closed loop controller andthe target open loop trajectory. Using this, we are able to have theOncilla perform a stable gait without requiring any memory, using anELM setup to generate the motor signals from the sensors. We showthat a stable closed loop limit cycle can be obtained using supervisedlearning for only a few periods of its gait, slowly transferring controlfrom the open to the closed loop.Finally, in chapter 7 we stretch the idea of morphological computation,and treat the whole legged robot with its controller as a single systemto be optimized. We optimize it using a deep neural network as controllerof a system by backpropagation through physics. To do this, wedeveloped a physics engine framework inside an automatic differentiationlibrary. This allows us to backpropagate through the controller,physics and renderer. We are able to show remarkably short optimizationprocesses despite only having quite complex sensory signals suchas cameras as inputs, in setups which are only partially observableand underactuated.We conclude that incorporating prior knowledge is beneficial whensetting up machine learning models for controlling robots. We alsoconclude that we were able to show that both morphological sensingand morphological control can be valid strategies for developingcontrollers for legged robots." @default.
• W2808875021 created "2018-06-29" @default.
• W2808875021 creator A5032189040 @default.
• W2808875021 date "2018-01-01" @default.
• W2808875021 modified "2023-09-24" @default.
• W2808875021 title "Incorporating prior knowledge into deep neural network controllers of legged robots" @default.
• W2808875021 hasPublicationYear "2018" @default.
• W2808875021 type Work @default.
• W2808875021 sameAs 2808875021 @default.
• W2808875021 citedByCount "0" @default.
• W2808875021 crossrefType "dissertation" @default.
• W2808875021 hasAuthorship W2808875021A5032189040 @default.
• W2808875021 hasConcept C104317684 @default.
• W2808875021 hasConcept C108583219 @default.
• W2808875021 hasConcept C11413529 @default.
• W2808875021 hasConcept C119857082 @default.
• W2808875021 hasConcept C151730666 @default.
• W2808875021 hasConcept C154945302 @default.
• W2808875021 hasConcept C185592680 @default.
• W2808875021 hasConcept C202444582 @default.
• W2808875021 hasConcept C2779343474 @default.
• W2808875021 hasConcept C33923547 @default.
• W2808875021 hasConcept C34413123 @default.
• W2808875021 hasConcept C41008148 @default.
• W2808875021 hasConcept C45374587 @default.
• W2808875021 hasConcept C50644808 @default.
• W2808875021 hasConcept C55493867 @default.
• W2808875021 hasConcept C66746571 @default.
• W2808875021 hasConcept C81363708 @default.
• W2808875021 hasConcept C86803240 @default.
• W2808875021 hasConcept C90509273 @default.
• W2808875021 hasConcept C92757383 @default.
• W2808875021 hasConceptScore W2808875021C104317684 @default.
• W2808875021 hasConceptScore W2808875021C108583219 @default.
• W2808875021 hasConceptScore W2808875021C11413529 @default.
• W2808875021 hasConceptScore W2808875021C119857082 @default.
• W2808875021 hasConceptScore W2808875021C151730666 @default.
• W2808875021 hasConceptScore W2808875021C154945302 @default.
• W2808875021 hasConceptScore W2808875021C185592680 @default.
• W2808875021 hasConceptScore W2808875021C202444582 @default.
• W2808875021 hasConceptScore W2808875021C2779343474 @default.
• W2808875021 hasConceptScore W2808875021C33923547 @default.
• W2808875021 hasConceptScore W2808875021C34413123 @default.
• W2808875021 hasConceptScore W2808875021C41008148 @default.
• W2808875021 hasConceptScore W2808875021C45374587 @default.
• W2808875021 hasConceptScore W2808875021C50644808 @default.
• W2808875021 hasConceptScore W2808875021C55493867 @default.
• W2808875021 hasConceptScore W2808875021C66746571 @default.
• W2808875021 hasConceptScore W2808875021C81363708 @default.
• W2808875021 hasConceptScore W2808875021C86803240 @default.
• W2808875021 hasConceptScore W2808875021C90509273 @default.
• W2808875021 hasConceptScore W2808875021C92757383 @default.
• W2808875021 hasLocation W28088750211 @default.
• W2808875021 hasOpenAccess W2808875021 @default.
• W2808875021 hasPrimaryLocation W28088750211 @default.
• W2808875021 hasRelatedWork W114497497 @default.
• W2808875021 hasRelatedWork W1516428299 @default.
• W2808875021 hasRelatedWork W2300668509 @default.
• W2808875021 hasRelatedWork W2547333951 @default.
• W2808875021 hasRelatedWork W2553534546 @default.
• W2808875021 hasRelatedWork W2603174016 @default.
• W2808875021 hasRelatedWork W2731869155 @default.
• W2808875021 hasRelatedWork W2771549791 @default.
• W2808875021 hasRelatedWork W2799517192 @default.
• W2808875021 hasRelatedWork W2898399448 @default.
• W2808875021 hasRelatedWork W2899498817 @default.
• W2808875021 hasRelatedWork W2953250718 @default.
• W2808875021 hasRelatedWork W2982252571 @default.
• W2808875021 hasRelatedWork W3017303660 @default.
• W2808875021 hasRelatedWork W3091224493 @default.
• W2808875021 hasRelatedWork W3120428840 @default.
• W2808875021 hasRelatedWork W3152883941 @default.
• W2808875021 hasRelatedWork W3187631620 @default.
• W2808875021 hasRelatedWork W3200465827 @default.
• W2808875021 hasRelatedWork W2468932347 @default.
• W2808875021 isParatext "false" @default.
• W2808875021 isRetracted "false" @default.
• W2808875021 magId "2808875021" @default.
• W2808875021 workType "dissertation" @default.