Deep Learning & Reinforcement Learning

While studying for my third year of university my cohort was tasked to develop a generative model and a reinforcement learning agent. The generative model should use subsets of the CIFAR10 (32x32) & STL10 (96x96) image datasets to develop pictures of a pegasus. As seen in my work below, I developed a GLO (generative latent optimisation) model, which produced good reproductions, but was poor for interpolating images.
While the reinforcement learning agent should adapt to play the Atari game Gravitar. The aim of Gravitar is complex, even for humans. The player must meander around gravity, visiting planets, collecting fuel, all while shooting other space ships. To develop my agent, I looked towards modern reinforcement learning recurrent network architectures as I thought this would perform well within the context of the game to satisfy a long term reward.
Please also see the video below which contains my RL agent's PB at Gravitar after training for 12 hours on a single GPU.
Libraries & Technologies used: PyTorch, CUDA, OpenAI Gym.