A reproducible hybrid approach to learning autonomous control of a robotic device has been proposed and experimentally verified, implemented using publicly available tools without specialized computing hardware. The approach combines learning from operator demonstrations, reinforcement learning, curriculum learning, and domain randomization into a single sequential framework. Based on a comparative analysis of five prototypes representing key classes of architectural solutions, applicable approaches for resource-constrained environments are identified. Three controlled experiments in Unity ML-Agents demonstrated the superiority of the hybrid combination over isolated approaches in terms of both learning speed and robustness to environmental changes. Limitations associated with catastrophic forgetting are outlined, and prospects for integrating large language models to generate synthetic training data are identified.