Sim-to-Real Multirotor Controller Single-shot Learning

Abstract

This paper demonstrates the sim-to-real capabilities of retrospective cost optimization-based adaptive control for multirotor stabilization and trajectory-tracking problems. First, a continuous-time version of the widely used discrete-time retrospective control adaptive control algorithm is developed. Next, a computationally inexpensive 12-degree-of-freedom model of a multirotor is used to learn the control system in a simulation environment with a single trajectory. Finally, the performance of the learned controller is verified in a complex and realistic multirotor model in simulation and with a physical quadcopter in a waypoint command and a helical trajectory command.