Fidelity Invariance Curriculum Learning : Accelerating Learning of Drone Navigation Policies using a General Multi-Fidelity Curriculum Learning Framework

Sammanfattning

This thesis introduces Fidelity Invariance Curriculum Learning (FICL), a general reinforcement learning framework that integrates curriculum learning with multi-fidelity optimisation to reduce training costs and improve policy generalisation. Building on Teacher-Student Curriculum Learning (TSCL), FICL incorporates cost-aware task sampling and leverages both low- and high-fidelity environments. To evaluate the approach, we construct a quad-rotor drone navigation scenario where an agent must locate and reach a target using GPS-like, approximate coordinates and visual input. FICL is compared against two baselines: high-fidelity TSCL and high-fidelity non-curriculum learning. Comparisons use convergence cost, episodic return, and episode termination metrics. Experimental results show that FICL achieves faster convergence in wall-clock time and maintains policy performance compared to baselines, while also contributing to more stable training dynamics, demonstrating improved sample efficiency without compromising solution quality. Results suggest that combining curriculum learning with fidelity-aware task sampling provides a promising approach for scalable reinforcement learning in resource-constrained domains, such as autonomous navigation.