Robotic manipulation policies are often trained with limited expert data, in complex actionspaces, and deployed under environmental variability. This thesis compares explicit BehaviorCloning (BC) and Diffusion Policy (DP) on benchmark tasks from the ManiSkill3 framework,and examines their performance. BC offers a lightweight, one-step regression approach but suffers from covariate shift and fails to capture multimodal action distributions. DP frames policylearning as a conditional denoising diffusion process, yielding robust, multimodal action trajectories at the cost of slower inference. DP consistently outperforms BC in sample efficiency, tasksuccess rate, and generalization to novel environments across varying demonstration set sizes(100–600) and two manipulation tasks (pick-and-place and cube stacking). However, DP’s inference speed (≈17 Hz) limits its use in high-frequency control, whereas BC achieves >1,700Hz. These results show the trade-off between accuracy and real-time feasibility.