Single-Shot Learning of Multirotor Controller Gains: A Data-Driven Approach with Experimental Validation

TL;DR

This paper presents a data-driven, single-shot learning method for multirotor controller gains using retrospective cost optimization, validated through simulations and real-world experiments, simplifying implementation and ensuring robust trajectory tracking.

Contribution

Introduces a continuous-time retrospective control-based approach for single-shot learning of multirotor gains, validated in simulation and real-world tests, improving simplicity and robustness.

Findings

Successful transfer of learned gains from simulation to physical quadrotor

Effective trajectory and waypoint tracking achieved in experiments

Performance independent of sampling time choice

Abstract

This paper demonstrates the single-shot learning capabilities of retrospective cost optimization based data-driven control applied to learning multirotor controller gains for trajectory tracking. In particular, the proposed control approach is first used within a simple multirotor simulation environment to learn appropriate multirotor controller gains to follow a trajectory. Then, the gains resulting from a single simulation run are used in a more complex multirotor simulation environment based on Simulink for performance verification. Finally, the resulting gains are implemented in a physical quadrotor and the results for waypoint and trajectory tracking are reported in this paper. The proposed control approach is the continuous-time version of the widely used discrete-time retrospective control adaptive control algorithm, which is simpler to implement within continuous-time simulation environments and whose performance does not depend on appropriate sampling time choice.