Data-driven exoskeletons promise adaptive augmentation of human mobility. Yet their widespread adoption is hindered by labor-intensive biomechanical data collection and extensive manual tuning. This study presents a highly efficient, simulation-generated synthetic data approach. It also designs a model-free algorithm for variable-speed walking to validate the method. We leveraged an Adversarial Motion Priors (AMP) agent to learn stylized walking within a massively parallel, physics-based simulation. The resulting high-fidelity data were collected and validated against OpenSim inverse dynamics pipelines. A novel CNN-Transformer architecture was developed to map contralateral swing-phase sensor data to variable-length push-off torque profiles. This enables real-time, adaptive torque assistance for exoskeletons. Experimental validation on a custom ankle exoskeleton demonstrated robust sim-to-real transferability. The system achieved approximately 85% torque prediction accuracy across speeds ranging from 0.6 to 1.75 m·s⁻¹. The controller significantly reduced user ankle positive mechanical work, thereby lowering metabolic demand. Furthermore, our multi-sensor configuration exhibited inherent fault tolerance, ensuring safe operation even under partial sensor failure. By replacing handcrafted control strategies with a scalable, data-driven approach, this work offers a practical pathway toward deploying autonomous exoskeletons in unconstrained, real-world environments.