This project attempts to replicate results obtained in the paper Time-Contrastive networks: self-supervised learning from video. We perform experiments to show that videos taken from a single viewpoint can be used to learn a distance preserving embedding of a scene onto a vector space. We attempt to teach a simulated robotic arm to imitate itself performing different trajectories using reinforcement learning and the learned embedding. The results show that an embedding can be learned in a self-supervised manner. However, learning to imitate proved itself more challenging.
More in depth discussion about our project can be found from our report!