A hybrid framework for UAV obstacle avoidance integrating reactive sensing, LiDAR planning, and deep reinforcement learning in real time

This work presents LT-DQN, a three-tier, safety-gated obstacle avoidance framework that combines a learned DQN controller, LiDAR-based A* planning, and a lightweight Three-Beam reactive fallback for UAV navigation using only onboard sensing. The hybrid method is evaluated in Unreal Engine with Microsoft AirSim across two environments with static and dynamic obstacles under both single-UAV and multi-UAV settings, and is compared against its individual component baselines. Across the tested configurations, LT-DQN outperforms these baselines by achieving a 100% success rate with zero collisions in both single-UAV and multi-UAV trials, while maintaining favorable safety and efficiency metrics, including a 5.17 ± 2.09 m minimum clearance and a 40.3 ± 6.0 s time-to-goal in the multi-UAV setting at 6 m/s. Real-time feasibility is supported by a low control-cycle time of 2.60 ± 0.31 ms, with occasional overhead during safety-layer switching. Overall, these findings indicate that LT-DQN achieves an effective balance between safety and efficiency, enabling robust UAV navigation in cluttered environments.