MobileNetV3-Small–Driven PQCN for Robotic Sorting in Clutter: Generalization Across Diverse Geometric Objects

This study evaluates a lightweight value-based deep reinforcement learning framework for multi-step robotic object sorting in cluttered scenes. We implement a Pixel-wise Q-learning Critic Network with a MobileNetV3-Small backbone and a fully convolutional value estimator (PQCN-MobileNetV3-Small–FCN-1152) to predict dense action-value maps for push, grasp, and place primitives from visual observations. In a simulated robotic sorting environment featuring diverse object geometries and clutter densities, we examine how the composition of the training object set affects policy generalization. Specifically, we compare the performance of PQCN models trained on regular (cubic) blocks with those trained on irregularly shaped blocks. We report performance using task success and action-efficiency metrics across multiple clutter regimes that emulate common sorting challenges such as occlusion and congestion. The results show that both training modalities enable effective multi-step sorting, while training on regular blocks yields higher robustness in very dense clutter, whereas training on irregular objects can improve performance in less congested settings. These findings highlight the role of training data design and lightweight perception backbones in improving the generalization and efficiency of deep reinforcement learning policies for robotic object sorting in simulation.