Physics-constrained multimodal reinforcement learning for local UAV Navigation in complex static obstacle environments

Abstract

Local navigation in complex static obstacle environments requires an unmanned aerial vehicle to reach a target while avoiding obstacles under limited local perception. U-shaped non-convex obstacles acting as local traps further increase the difficulty of detour selection and continuous-control execution. A key challenge is to jointly capture forward scene structure, surrounding geometric constraints, and execution-level control requirements within a unified decision-making pipeline. To address this challenge, we present PC-MM-SAC, a task-oriented multimodal reinforcement-learning framework with a lightweight physics-constrained action-execution layer for local navigation in complex static obstacle environments. The method learns from structured observations built from depth maps, LiDAR measurements, and task-related state variables. The layer applies bounded action mapping and yaw-rate variation limiting to improve the smoothness and continuity of continuous-control commands without introducing an additional online optimization-based controller. We further employ an event-conditioned modality value-sensitivity analysis during evaluation to characterize how the critic’s local sensitivity to different information sources varies across representative navigation phases. In AirSim experiments, PC-MM-SAC achieves a success rate of 0.88 and an episode reward of 772, attaining the highest success rate among the evaluated methods under the current task setting. Ablation and behavioral analyses suggest that multimodal observations are the primary contributor to the observed performance improvement, while the physics-constrained action-execution layer is associated with smoother trajectory execution and reduced local oscillation.