Robotic assembly relies on the precise joining, fastening, and positioning of components. While classical motion-planning techniques excel in tightly structured workspaces requiring millimetric precision, they lack the flexibility to adapt to unstructured failures, often necessitating manual intervention or rigid programmed reactions. Deep Reinforcement Learning (DRL) offers a paradigm shift, allowing agents to learn optimal decisions through trial-and-error without explicit programming. Although DRL has demonstrated success in simpler or looser assembly benchmarks, learned policies frequently fall short of the positional accuracy required for strict industrial tolerances, often exhibiting stochastic behavior that is incompatible with high-precision manufacturing. The research specifically targets the Gear Meshing task, a contact-rich scenario selected because it demands exceptional precision, far exceeding the requirements of standard pick-and-place baselines.