The construction industry faces a shortage of workers, while earthmoving tasks remain difficult to automate because they require both tracking the evolving terrain and handling soil dynamics as the machine interacts with the material. This thesis studies how a reinforcement-learning pipeline can be structured to address this problem in Newton simulation with particle-based terrain representation, focusing on loose-soil surface manipulation in a setup where the shovel interacts with the soil as a kinematic collider. The policies are trained with Proximal Policy Optimization (PPO), with the largest runs using up to 2048 environments in parallel. A generic current-terrain / target-terrain framework is introduced and evaluated first on a Flattening Task with a flat, volume-consistent target, and then on a broader General Shaping Task with plane and smooth sinusoidal target surfaces.