Abstract <p>Robotic drug grasping in complex smart pharmacy environments requires accurate detection and efficient execution, yet existing methods struggle with clutter, overlapping, and small target objects. To improve robotic grasping of chaotic and variably shaped drugs, we propose a novel multi-stage framework that integrates enhanced perception, grasp detection, and trajectory planning. First, images are preprocessed with an improved Super-Resolution Convolutional Neural Network (SRCNN) to enhance input quality. Next, drug segmentation is performed using our YOLO-EASB instance segmentation algorithm (YOLOv5+E-A-SPPFCSPC+BIFPNC), and the most suitable targets are identified by evaluating mask completeness. The segmented drugs are then processed by our improved Adaptive Feature Fusion and Grasp-Aware Network (IAFFGA-Net) with an optimized loss function, ensuring robust grasp detection in cluttered environments. For execution, we combine improved Particle Swarm Optimization and 3-5-3 interpolation to generate efficient and smooth robotic arm movements. Finally, the system is deployed on an adaptive collaborative robot capable of adjusting to diverse production environments. Experiments on our custom dataset demonstrate 97.3% drug recognition accuracy, while chaotic drug grasping tests achieved at least 80% success, satisfying the requirements of intelligent pharmacies.</p> Graphical abstract <p>The framework of the multi-stage grasping network combined with an adaptive robotics mechanism</p>