End-to-end Reinforcement Learning for Time Optimal Quadcopter Flight
We introduce end-to-end reinforcement learning for optimal quadcopter flight. Our network directly outputs motor commands and is compared to a traditional method where the network interfaces with an INDI controller. In simulation our method excels due to its direct control authority. In real-world tests, although the performance gap narrows, our method maintains a slight advantage, demonstrating the potential of end-to-end algorithms.