Speaker
Description
The increasing popularity of Unmanned Aerial Vehicles (UAVs) across multiple application domains necessitates robust autonomous landing capabilities on diverse and challenging platforms. It also requires coping with environmental disturbances such as wind gusts, ground effect, and aerodynamic drag.
Conventional controllers such as PID and linear MPC are effective for flat-surface landings but struggle with the nonlinear, underactuated dynamics present in inclined scenarios.
This paper introduces a fully end-to-end Deep Reinforcement Learning (DRL) framework that directly maps kinematic sensor observations to motor commands for quadrotor landing on sloped platforms. The approach formulates the problem as a Markov Decision Process (MDP) and incorporates comprehensive aerodynamic disturbance modeling, curriculum learning, and reward shaping. By eliminating intermediate control abstractions, the controller enables direct optimization of the entire control pipeline using Proximal Policy Optimization (PPO).
Simulation results demonstrate competitive performance across slope angles from 0° to 30°, with the controller generalizing to unseen slopes and wind conditions without requiring separate low-level stabilization. Comparative analysis shows that while hybrid DRL-PID approaches offer strong robustness, the end-to-end method is highly competitive and effective.
To our knowledge, this is the first work to explore an end-to-end DRL controller for UAV landing on sloped platforms under environmental disturbances. The open-source implementation is available online to support reproducibility and future research.