Learning to Fly: Computational Controller Design for Hybrid UAVs with Reinforcement Learning

a review of these combined advantages of multicopters in fixed-wing planes a typical hybrid UAV can switch between copter and plane modes the copter mode gives the hybrid UAV more maneuverability while the plane mode is more efficient for long-distance flights controlling the hybrid UAV however is very challenging due to the complexity of the dynamics when both rotors and wings are active traditionally people design hybrid UAVs controllers manually the controller design process is highly specific to the dynamic model of each view AV in the design knowledge cannot easily be shared in this paper we propose an automatic way to design the controller for various hybrid UAVs using reinforcement learning the first step is geometry design we build a data set of premature eyes components for hybrid UAV designs users can create new vehicles by mixing and matching parts from this collection this allows us to create vastly different configurations of hybrid UAVs once we have a design we train a neural network controller using reinforcement learning to narrow the reality gap we propose a novel reward function and new integral block in the neural network controller here we throw the controller after training for 25 50 100 and 400 iterations without any knowledge of the specific flight modes in the transition state it manages to learn a functional velocity tracking controller after 400 iterations the controller allows for a smooth transition from loitering to advancing next we show our quad plain following a triangular trajectory using our trained neural network controllers you in this example ATL sitter is instructed to track a time-varying target velocity as closely as possible here the x-wing model is doing an autonomous flight by tracking a pre-programmed target velocity that forms a circle finally the double wing model is flying along an s-shaped trajectory to avoid obstacles the simulation experiments show that our model agnostic method is able to generate controllers for wildly different configurations of hybrid UAVs this video shows the ablation test with two key components of our pipeline showing how they dramatically lengthen the flight time and decrease the velocity tracking error once we trained the controller we deployed it on our where to verify is robustness because of the reality gap a naive controller without our proposed modifications crashes catastrophic ly when implemented in real hardware after incorporating the integral block and orientation cost proposed in our work our trained control was capable of finishing the flight task successfully our controller design generalizes to hybrid UAVs with significantly different dynamics for example in order to enter gliding mode after hovering the quad plane controller simply redistributes thrust between front and other rotors without needing to change its orientation too much while the X plane has to lean forward by a lot to transition from copter mode to gliding mode when this cloud plane is hovering it can be seen from the log that all five rotors are actively being used however when the cloud plane enters gliding mode we can infer from the large margin between the weight and the net thrust that the wings contributed some lift to balance the weight thank you for watching

Leave a Reply

Your email address will not be published. Required fields are marked *