A framework for vision-based robotic grasping and sorting in cluttered environments

Purpose The purpose of this study is to tackle the challenges associated with robotic gripping technology in disordered and cluttered environments, which are common in real-world applications. The goal is to develop a novel framework that empowers robots to efficiently and accurately grasp and classify objects in highly chaotic settings, thereby improving the practical applicability and versatility of robotic systems in complex, dynamic environments. Design/methodology/approach To achieve the research objectives, the authors designed a framework that combines deep learning techniques with off-policy reinforcement learning. The framework was trained in a simulated environment to generate a diverse set of images, which enabled accurate object detection and plane segmentation. An off-policy reinforcement learning approach was then used to identify the optimal grasping strategy, focusing on adaptability and precision in unpredictable settings. The practical applicability of this framework was demonstrated through the example of flexible pharmaceutical sorting. Findings The experimental results demonstrate the effectiveness and robustness of the proposed framework. The robotic arm achieved an impressive success rate of 97.50%, with an average processing time of 7.92 s per medicine box. This performance surpasses that of human workers, who achieved a success rate of 95.18% and experienced a noticeable decline in efficiency after 30 min of continuous work. In addition, compared to existing frameworks, this framework demonstrates a higher success rate in the medication box sorting task and improves efficiency by 6% in terms of average processing time. Originality/value The originality of this study lies in the development of a novel framework that effectively overcomes the limitations of existing robotic gripping technologies in unstructured and cluttered environments. Unlike traditional approaches, this framework integrates deep learning and off-policy reinforcement learning to optimize object handling in chaotic settings.