In order to address the issues of insufficient accuracy and generalization in traditional reinforcement learning methods for insertion assembly tasks, a robotic arm insertion assembly skill learning and compliant control strategy based on an optimized ELM (extreme learning machine) was proposed. Firstly, GSA-ELM (gravitational search algorithm-optimized extreme learning machine) model was established for robotic arm insertion skill learning, utilizing time-series datasets of contact forces, positions, and velocities from robotic arm demonstration insertion operations to optimize the initialization parameters of the ELM, thereby enhancing accuracy and stability of the model. Secondly, an adaptive force-compensation compliant control method for robotic arms was developed based on contact force-position prediction information, integrated with a position-velocity dual-loop control strategy to achieve real-time force-position feedback compensation for desired positions and operational speeds during insertion assembly. The results show that this model effectively ensures compliance and stability in robotic arm insertion operations. Through USB (universal serial bus) insertion experiments conducted on a UR3 robotic arm, it is evident that the proposed approach demonstrates practical feasibility and precision in real-world assembly scenarios, which provides a novel technical pathway for robotic arm insertion assembly tasks.