This thesis investigates the use of reinforcement learning to control a fourteen-joint anthropomorphic prosthetic hand performing a cylindrical power grasp in a Unity-based simulation environment. Two reinforcement learning approaches are compared: Proximal Policy Optimisation augmented with Generative Adversarial Imitation Learning and Behavioural Cloning (PPO+GAIL+BC), and pure Soft Actor-Critic (SAC) without any imitation signal. A custom reward function and a four-lesson curriculum were designed to progressively train the agent from basic two-finger contact to a full anatomical opposition grasp with a 2.0-second hold requirement. The PPO+GAIL+BC trainer achieved a sustained grasp success rate above 95% and completed all four curriculum lessons within one million environment steps on a single laptop, while the SAC trainer failed to advance beyond the easiest level and converged on a local optimum. The results demonstrate that a small number of expert demonstrations combined with on-policy reinforcement learning can solve high-dimensional prosthetic grasping tasks on resource-constrained hardware where pure off-policy methods fail. Directions for future work include multi-seed evaluation, sim-to-real transfer, and extending the curriculum to multiple grasp types and object geometries.