Individuals with upper limb amputations or congenital defects face substantial challenges in performing daily tasks, often exacerbated by limited functionality in the existing bionic prostheses. To address this, we present an AI-powered virtual reality (VR) system designed to train patients to control bionic prostheses with a high number of degrees of freedom using electromyographic (EMG) activity decoding. Current prosthetic control systems lack sophistication, relying on discrete muscle activity decoding and manual mode switching. Our system employs a novel neural network architecture pre-trained on healthy subjects, facilitating continuous decoding of residual muscle activity to control a virtual limb in the VR. Through three months of training, participants with congenital upper limb dysplasia achieved precise finger-specific control, demonstrating a low error rate and strong correlation between the target and the decoded kinematics. Importantly, the system’s adaptability and its user-friendly interface positively impacted patient experience, offering potential reduction of muscle atrophy risks and psychological benefits while preparing individuals for the use of an advanced prosthetic device. This innovative approach represents a significant step toward enhancing prosthetic functionality and improving the quality of life for upper limb amputees and congenital defect patients.