UAVs' perception and planning in cluttered dynamic environments

Abstract

The evolution of robotics has been consistently guided by an immutable foundational objective: the strategic substitution of human labor with autonomous systems in environments characterized by elevated risk, extreme physical demands, or operational impracticalities. This paradigm shift aims not only to mitigate occupational hazards but also to enhance operational efficiency and reliability in inherently perilous domains. Concurrently, Unmanned Aerial Vehicles (UAVs) have emerged as a transformative technology within this framework, demonstrating rapidly expanding deployment capabilities. Their operational envelope now robustly encompasses low-altitude environments and extends progressively into complex, unstructured scenarios—including dense urban settings, confined industrial facilities, and disaster-stricken regions—where traditional robotic platforms face significant limitations. This proliferation signifies a critical advancement in overcoming previously insurmountable environmental and logistical constraints, thereby actualizing the long-envisioned potential of robotic systems to assume critical roles in high-stakes applications. Under this background, the dynamic perception and planning of UAVs becomes vital significant. This thesis addresses critical challenges related to the efficient autonomous system of UAVs' flight in dynamic, cluttered environments. The main contributions of this thesis are summarized as follows:

Firstly, the thesis addresses the challenge of perceiving and avoiding multiple small, fast-moving objects for quadrotors equipped with only a low-cost RGB-D camera. We propose the 3D-SORT (Simple Online and Real-time Tracking in 3D Space) algorithm to enable fast and effective multi-object tracking. A hierarchical trajectory generation framework, integrating kinodynamic path searching with gradient-based optimization, is developed to facilitate collision avoidance maneuvers.

Secondly, the thesis introduces the FAPP (Fast and Adaptive Perception and Planning) system, enabling real-time navigation for UAVs in densely cluttered dynamic environments. The contribution comprises a novel point cloud segmentation methodology optimized for computational efficiency, a covariance adaptation technique resolving state estimation challenges for objects exhibiting divergent motion characteristics, a trajectory generation framework incorporating uncertainty quantification for dynamic collision avoidance, and an adaptive re-planning contingency mechanism for optimization failure conditions.

Then, the thesis proposes a learning methodology for dynamic obstacle avoidance. A deep reinforcement learning framework is proposed, explicitly incorporating both static and dynamic obstacles during training. An end-to-end deep neural network is trained to extract kinematic features of all obstacles directly from environmental maps, subsequently generating quadrotor acceleration commands to ensure reliable collision avoidance.

Finally, the thesis addresses the UAVs' obstacle avoidance in high-speed flight. The proposed framework comprises three components: a novel incremental robocentric mapping technique utilizing distance and gradient data, an obstacle-aware topological path search algorithm capable of producing diverse feasible paths, and an adaptive gradient-based trajectory generation method incorporating a novel time pre-allocation strategy, optimized for high-velocity navigation.