A Reinforcement Learning Approach to Control of a Quadrotor Biplane Tailsitter for Adaptive Landing Maneuvers

Presented at the Vertical Flight Society 79th Annual Forum & Technology Display
System Engineering Tools and Processes Technical Session - Paper 1208
11 pages

https://doi.org/10.4050/F-0079-2023-18173

Abstract:
This paper presents a reinforcement learning (RL) based trajectory planning and control architecture for autonomous landing maneuvers of a quadrotor biplane tailsitter (QRBP). The RL controller replaces a gradient-descent based optimal trajectory planner and outer loop position controller of a standard QRBP control system, while retaining the inner loop for regulating attitude dynamics. The RL agents are trained in a simulated environment, using a curriculum learning approach for training an RL agent capable of landing on a moving Unmanned Ground Vehicle (UGV) with changing velocity. The RL architecture is capable of generating landing trajectories onto a moving ground vehicle, with computational costs suitable for real-time implementation. Further, the RL guidance architecture successfully completes the landing mission more consistently compared to a gradient-descent based guidance architecture when there is uncertainty in the path of the UGV.

Did you attend Forum 79? Click the preview below to access the full paper.

The full paper is also available for purchase at the Vertical Flight Society Online Library and Store.