The recent advancements in visuomotor control for robotics have shown a shift towards more task-specific and hierarchical object representations, which enhance the efficiency and robustness of learning policies. Innovations in object-centric approaches have demonstrated significant improvements in sample efficiency and generalization, particularly in long-horizon tasks. These approaches leverage hierarchical decompositions of scenes and objects, enabling selective representation assembly tailored to specific tasks. Additionally, there is a growing focus on few-shot learning and intra-category transfer, which allows robots to learn complex tasks from minimal demonstrations by optimizing object arrangements in canonical frames. Another notable trend is the integration of advanced optical sensing technologies into surgical robotics, with models like Memorized Action Chunking with Transformers (MACT) showing promise in automating tissue surface scanning through efficient imitation learning. Lastly, the field is exploring alternative motion strategies such as pick-and-toss (PT) to enhance task efficiency, with methods that dynamically choose between PT and pick-and-place (PP) based on estimated task difficulty.