Traditional learning approaches proposed for controlling quadrotors orhelicopters have focused on improving performance for specific trajectories byiteratively improving upon a nominal controller, for example learning fromdemonstrations, iterative learning, and reinforcement learning. In theseschemes, however, it is not clear how the information gathered from thetraining trajectories can be used to synthesize controllers for more generaltrajectories. Recently, the efficacy of deep learning in inferring helicopterdynamics has been shown. Motivated by the generalization capability of deeplearning, this paper investigates whether a neural network based dynamics modelcan be employed to synthesize control for trajectories different than thoseused for training. To test this, we learn a quadrotor dynamics model using onlytranslational and only rotational training trajectories, each of which can becontrolled independently, and then use it to simultaneously control the yaw andposition of a quadrotor, which is non-trivial because of nonlinear couplingsbetween the two motions. We validate our approach in experiments on a quadrotortestbed.
展开▼
