Dış Kaynaklı Proje Süreçleri Yönetim Sistemi
In this project, we realize mechanisms for the robots that enable them to discover both high-level low-dimensional discrete object and action primitives and their associated operators, and to learn low-level high-dimensional continuous predictive models that are used in both high and low-level planning and verification. To the best of our knowledge, simultaneous discovery of object, effect, and action categories, learning their associated operators, and using these operators for planning has not been studied in the literature yet. Different deep neural network architectures are proposed for the discovery of action primitives. Firstly, we suggest passing the sensorimotor trajectory into a conditional neural process network, taking the sensorimotor trajectory as input and extracting the discrete category of the motor component through a discrete bottleneck layer. Alternatively, using the same network structure, we propose to create both continuous and discrete latent representations, and suggest using discrete representation to filter continuous data, thus allowing categories to emerge in the discrete representation. Finally, we propose a structure that obtains both object and action categories using Attention architecture and discrete activation layers.