End-to-end Neural Network Based Optimal Quadcopter Control
We present for the first time, a flight tested end-to-end Guidance & Control Network for quadcopters that does not rely on innerloop controllers for stabilization, but instead gives direct motor commands. This approach gives full control authority to the network, but it is more sensitive to modeling errors. To mitigate this challenge we train an "adaptive" G&CNet that can find the optimal motor command for a given moment model mismatch. By measuring and incorporating this model mismatch in real-time, we achieve improved performance. Even when subjected to significant disturbances such as the addition of external weight, our controller seamlessly adapts and automatically discovers the optimal trajectory!
For more detailed information, please refer to our paper: https://www.sciencedirect.com/science...