Robotic manipulation in unstructured environments, a key pillar of modern automation, relies heavily on advanced visual perception. This paper provides a comprehensive survey of recent progress in vision-based robotics, with a focus on the challenges of 6D pose estimation and grasp detection in robotic bin-picking tasks. It traces the transition from traditional geometry-based approaches to deep learning-driven methods, emphasizing innovations such as hierarchical surface encoding, self-occlusion reasoning, multi-view fusion, and temporal modeling. The study also reviews grasp detection techniques that produce dense, stability-aware predictions, allowing robots to perform effectively in cluttered scenarios. Addressing the “data dilemma”—the high cost of manual annotation—the paper examines emerging solutions like synthetic datasets, sim-to-real transfer, and self-supervised learning. Finally, it explores advances in generalization, where few-shot and zero-shot learning enable robots to manipulate unfamiliar objects. Overall, this survey offers an integrated perspective on developing data-efficient, resilient, and generalizable robotic perception systems.