Key point detection of lightweight robotic arms based on transfer learning

Traditional robotic arm pose estimation methods predominantly rely on high-precision sensors. However, these approaches often encounter challenges such as elevated costs and complex installation procedures in practical scenarios. In contrast, vision-based pose estimation leverages visual sensors, such as cameras, to capture the state of the robotic arm. By utilizing deep learning techniques to analyze the acquired image data, keypoint information is extracted and subsequently transformed into control commands. This approach offers advantages in terms of simplicity and ease of maintenance. Nonetheless, vision-based methods generally require substantial datasets and pose significant challenges when applied in environments with constrained computational resources. This study introduces a novel robotic arm keypoint detection network, which combines transfer learning with an enhanced YOLOv8 algorithm. The COCO human pose dataset is employed as the source dataset for pre-training, enabling the extraction of generalized features. Following this, fine-tuning is conducted using a specific robotic arm dataset. Experimental results demonstrate that the proposed method, after undergoing transfer learning, achieves a 3.8% increase in mAP50-95. Additionally, it significantly reduces the number of parameters and floating-point operations, thereby making it more suitable for deployment in environments with limited computational capacity.