Autonomous navigation in unstructured outdoor environments using semantic segmentation guided reinforcement learning

Abstract

Robust autonomous navigation in dense, unstructured environments such as forests presents a longstanding challenge in robotics due to complex terrain geometry, dynamic occlusions, and unreliable global positioning signals. This paper proposes a hybrid perception-and-control framework that integrates deep semantic segmentation with reinforcement learning to enable intelligent, vision-driven navigation in visually cluttered forest trails. The system combines Mask R-CNN for pixel-level trail segmentation with a Soft Actor-Critic (SAC) agent that learns adaptive navigation policies under continuous action spaces. A Pure Pursuit controller translates visual predictions into smooth motor commands, ensuring path adherence and stability. The model is trained and evaluated in a high-fidelity forest simulation environment featuring natural obstacles, variable lighting, and randomized trail geometries. Extensive experiments demonstrate that our approach achieves a high trail-following success rate (86.7%), low collision frequency, and precise path tracking in challenging navigation scenarios. Comparative and ablation studies further highlight the synergy between learning-based perception and control. The proposed framework offers a scalable and modular solution for deploying autonomous robots in natural terrains without relying on GPS or prior maps, paving the way for applications in environmental monitoring and field robotics.