This thesis explores reinforcement learning in a Unity environment, moving step by step from simple energy collection to complex obstacle avoidance. It uses Unity ML‑Agents to train agents and tests how reward design affects learning. The project shows that small mistakes in rewards can cause unusual behaviors, like the agent freezing or crashing on purpose. Through iterative tuning and debugging, high‑performing agents with high success rates across all levels was achieved. Overall, the thesis showcases strong technical execution, clear documentation of challenges, and valuable insights into designing adaptive and interpretable RL training setups.