Adaptive-Risk-Aware End-to-End Control System for Mapless Navigation of Drones

Safe and autonomous navigation of drones in 3D complex unknown environments remains a current research hotspot. A critical challenge is how to adaptively balance the safety with flight aggressiveness, fully exploiting the motion potential of the drone. This letter proposes an Adaptive-Risk-Aware End-to-End (ARA-E2E) control system which can generate real-time control commands capable of achieving adaptive balance in 3D environments. The system utilizes the Deterministic Implicit Quantile Network (DIQN) algorithm under the Actor-Critic framework, directly generating acceleration commands from LiDAR data. In the critic network, the Implicit Quantile Networks (IQN) algorithm, which integrates a risk-aware module, is employed. IQN evaluates actions with varying tendencies toward flight speed and risk, by calculating the parameter of Conditional Value-at-Risk (CVaR) based on the current obstacle environment. In the actor network, a continuous state space representation method is designed to provide more flexible control commands. The comparative experiments show that the proposed method outperforms in flight time, success rate, and energy consumption in the forest scenario. To further validate the effectiveness of the system, we deploy it on an autonomous drone platform, and flight tests are conducted in an outdoor forest environment. The drone can successfully complete the real-world navigation task at an average speed of 2.5 m/s (with a maximum of 5 m/s in the simulation), demonstrating the practicality and reliability of the proposed method.