Despite rapid advances in robot learning, the coordination of closed-loop manipulation in cluttered environments remains a challenging and relatively underexplored problem. We present a novel two-level hierarchical architecture for a depth vision-equipped robotic arm that integrates pushing, grasping, and high-level decision making. Central to our approach is a prioritised action–selection mechanism that facilitates efficient early-stage learning via behaviour cloning (BC), while enabling scalable exploration through reinforcement learning (RL). A high-level decision neural network (DNN) selects between grasping and pushing actions, and two low-level action neural networks (ANNs) execute the selected primitive. The DNN is trained with RL, while the ANNs follow a hybrid learning scheme combining BC and RL. Notably, we introduce an automated demonstration generator based on oriented bounding boxes, eliminating the need for manual data collection and enabling precise, reproducible BC training signals. We evaluate our method on a challenging manipulation task involving five closely packed cubic objects. Our system achieves a completion rate (CR) of 100%, an average grasping success (AGS) of 93.1% per completion, and only 7.8 average decisions taken for completion (DTC). Comparative analysis against three baselines—a grasping-only policy, a fixed grasp-then-push sequence, and a cloned demonstration policy—highlights the necessity of dynamic decision making and the efficiency of our hierarchical design. In particular, the baselines yield lower AGS (86.6%) and higher DTC (10.6 and 11.4) scores, underscoring the advantages of content-aware, closed-loop control. These results demonstrate that our architecture supports robust, adaptive manipulation and scalable learning, offering a promising direction for autonomous skill coordination in complex environments.
Yu et al. (2025) studied this question.