Dynamic Ego-Centric Graph-Based Reinforcement Learning for Autonomous Quadrotor Navigation

Efficient and safe autonomous navigation of quadrotors in cluttered and partially observable environments remains a challenging problem due to the constraints imposed by quadrotor dynamics, high-dimensional sensory inputs, and complex obstacle configurations. This paper proposes a hybrid learning framework that integrates Graph Neural Networks (GNNs) with Deep Reinforcement Learning (DRL) to enable adaptive and collision-free quadrotor navigation. The environment is represented as a dynamic, ego-centric graph, where nodes encode local spatial regions or obstacles and edges capture traversability relationships. A GNN-based encoder extracts structured, context-aware embeddings from this representation, which are fused with the quadrotor's dynamic state and provided as input to a Proximal Policy Optimization (PPO) agent for continuous control. The proposed framework is evaluated in a PyBullet simulation environment under identical conditions against standard PPO baselines using flat, unstructured vector inputs. Experimental results demonstrate that incorporating graph-based environmental reasoning leads to substantial and consistent improvements in navigation performance, including a 27% increase in success rate (from 65% to 92%), higher cumulative rewards, smoother trajectories, and a 65% reduction in reward variance. These quantitative gains, coupled with a 45% improvement in obstacle clearance, highlight the effectiveness of structured relational representations in enhancing the robustness, efficiency, and stability of learning-based aerial navigation policies.