Adaptive Quadrotor Control

Unmanned aerial vehicles (UAVs) play an increasing role in a wide number of missions such as remote sensing, transportation, and search and rescue missions. Often, a critical part of a UAV's role is to carry loads vital to the mission. For example, cargoes may consist of food and supply delivery in disaster struck areas, patient transport, or spacecraft landing. Planning motions for a UAV with a load becomes difficult because load swing is difficult to control. However, it is a necessity for the safety and success of the mission.

We are interested in producing a trajectory for an aerial robot with a suspended load that delivers the load to a destination in a swing-free fashion while avoiding static obstacles.

We developed a motion planning framework for generating trajectories with minimal residual oscillations (swing-free). We rely on reinforcement learning algorithms to train the system for a particular load and to transfer the learned policy to different models and state and action spaces. Using this methodology, we are able to train the agent in two dimensional action space and fly a quadrotor in three dimensions, including ascending and descending. Integration of the Reinforcement learning with Probabilistic Roadmaps (PRMs) and path tracking algorithm allows us to create a trajectory with bounded load displacement that automatically avoids static obstacles.