Toward Embodied Intelligence: An Autonomous Exploration System for Multimodal UAV

Unmanned Aerial Vehicles (UAVs) embodied intelligence is of increasingly important since UAVs are well suited to perform tasks that are dangerous or impractical for humans. Numerous model-based approaches for autonomous UAV exploration have been proposed, yet they predominantly rely on a single sensor. This reliance limits the range of information obtained and diminishes the success rate in search missions. In this paper, in order to tackle those issues, a multimodal UAV autonomous exploration system is proposed. A pioneering method is introduced that harnesses deep reinforcement learning to synergize spectral data with LiDAR information, thereby significantly enhancing the navigation capabilities of UAVs in complex environments. A novel UAV equipped with spectral sensing and conducted extensive experiments is developed on a virtual simulation platform. Simulation results demonstrate that the proposed method demonstrating increased robustness and a higher success rate in obstacle avoidance.