The article presents a new approach to controlling quadcopters for time-optimal flight using end-to-end reinforcement learning (E2E). Traditional quadcopter control methods face difficulties with unmodeled effects. Recent advancements have brought reinforcement learning (RL) to the forefront, showcasing its potential in achieving high-speed flight. However, the reality gap between simulation and real-world application poses a challenge. The authors’ novel E2E approach addresses this by giving direct motor commands, incorporating a learned residual model, and an adaptive method to compensate for modeling errors. The E2E approach showed a significant advantage over the state-of-the-art network in both simulation and real-world testing, highlighting its potential.
Publication date: 29 Nov 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2311.16948
