I have been working with the implementation of a mobile robot control paper, which guides an agent in a dynamic environment with unknown trajectories. I mostly worked by the paper’s explanations: paper, but my model doesn’t seem to converge. I trained both a DQN and a PPO agent for around 100 thousand episodes, and debugged potential issues for quite a while now but nothing yielded “good-like” results. I’m a second year BSc student and would say I’m only exploring reinforcement learning concepts, so any advice would be well appreciated. I’m generally asking about implementation structure and initializations, do I make a rookie mistake in the base code or did I misunderstand something? I’m generally interested in my maskPPO agent’s and my CNN model’s correctness, are they implemented correctly..? My git repo is reachable here with the full project: repo