Deep learning: creating bridges between DMPs in autoencoders and recurrent neural networks

The complexity in modeling human movement increases as the dimensionality of these movement grows. Since searching more precision and flexibility involves more variables in the model. Dynamic Movement Primitives (DMP) have shown the ability to generate joint movements with high complexity. However,...

Descripción completa

Detalles Bibliográficos
Autor: Arbones Clua, Marc
Tipo de recurso: tesis de maestría
Fecha de publicación:2017
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/109222
Acceso en línea:https://hdl.handle.net/2117/109222
Access Level:acceso abierto
Palabra clave:Neural networks (Computer science)
Robotics
Xarxes neuronals (Informàtica)
Robòtica
Àrees temàtiques de la UPC::Informàtica
Descripción
Sumario:The complexity in modeling human movement increases as the dimensionality of these movement grows. Since searching more precision and flexibility involves more variables in the model. Dynamic Movement Primitives (DMP) have shown the ability to generate joint movements with high complexity. However, the problem remains in the interaction between several joints since DMP alone is not able to deal with it. To solve this problem a new model called autoencoded dynamic movement primitive (AE- DMP) is introduced in the work "Efficient movement representation by embedding DynamicMovement Primitives in Deep Autoencoders"[2]. The proposed approach uses autoencoder in order to find a representation of the movement in a latent space. Consequently, the DMPmodel is able to reconstruct the complete movement. In thisMaster Thesis we will study the implementation of this model and study its performance. All the features stated in the original paper are checked, as multiple movements, sparsity and reconstruction of missing or corrupted data.