Lightweight Safe Reinforcement Learning for End-to-End UAV Navigation

This paper presents a safety-constrained perception-control framework for end-to-end UAV navigation. A lightweight encoder uses asymmetric and depthwise separable convolutions to extract collision-risk-aware features from sparse observations. The navigation task is formulated as a constrained Markov decision process and optimized with a Lagrangian safe PPO algorithm within a hierarchical controller. Curriculum learning is used to improve training stability. The abstract reports higher success rates, better safety, and improved efficiency than reinforcement-learning baselines across obstacle densities and flight speeds, but provides no numerical results, benchmark details, or deployment measurements.