A robotic arm control system integrating particle swarm optimization-genetic cooperative optimization (PSO-GA) and adaptive fuzzy proportional-derivative (AF-PD) control was proposed to address challenges in dynamic gesture recognition based on surface electromyography (sEMG), including time variability, nonlinearity, and difficulties in traditional control parameter tuning. Eight-channel sEMG signals were first collected, and multi-dimensional features were extracted to establish an end-to-end control architecture. The PSO-GA co-optimization algorithm was proposed to parallelly optimize the path constraint parameters of dynamic time warping (DTW) and the hyperparameters of the random forest (RF) classifier. An average recognition accuracy of 98.7%±0.6% is achieved on a self-constructed six-gesture dataset, which is 5.2% higher than that of the traditional RF-DTW method, while training time is reduced by 40%. A recognition rate of over 94.5%±1.0% is maintained on the public Ninapro DB1 dataset, demonstrating strong generalization capability. An AF-PD controller is designed, integrating a gesture library and a joint mapping strategy to enable smooth transitions from discrete gestures to continuous joint motion. Experimental validation shows that the system achieved a control delay of less than 100 ms and a task operation success rate of 96.3%. This work is of positive significance for advancing rehabilitation aids and human-computer interaction technologies.