This project aims to create a modular Autoencoder training and inference pipeline. Different architectures of Autoencoders can easily be added for image representation generation along with image reconstruction.
The pipeline is designed to read a config file to dynamically generate the relavant input, output layer dimensions along with bottleneck layer size. Other training related hyperparameters are also read from the config file.
The dataset used for testing consists of images captured from the wrist camera of a Kinova3 robot arm inside a coppeliaSim environment.
🎓 This codebase is part of the authors Master Thesis titled Visuomotor Policy Learning for Predictive Manipulation
Anirudh NJ
Email: [email protected]
GitHub: @njanirudh