MRAC-RL: A Framework for On-Line Policy Adaptation Under Parametric Model Uncertainty

TL;DR

This paper introduces MRAC-RL, a novel framework combining adaptive control and reinforcement learning to enable policies trained in simulation to adapt effectively to real-world systems with parametric uncertainties.

Contribution

The paper develops a set of MRAC algorithms applicable to linear and nonlinear systems, integrating adaptive control with RL for improved real-world deployment.

Findings

MRAC-RL enhances policy robustness against model uncertainties

The framework outperforms existing RL methods in uncertain environments

Demonstrated effectiveness on both linear and nonlinear systems

Abstract

Reinforcement learning (RL) algorithms have been successfully used to develop control policies for dynamical systems. For many such systems, these policies are trained in a simulated environment. Due to discrepancies between the simulated model and the true system dynamics, RL trained policies often fail to generalize and adapt appropriately when deployed in the real-world environment. Current research in bridging this sim-to-real gap has largely focused on improvements in simulation design and on the development of improved and specialized RL algorithms for robust control policy generation. In this paper we apply principles from adaptive control and system identification to develop the model-reference adaptive control & reinforcement learning (MRAC-RL) framework. We propose a set of novel MRAC algorithms applicable to a broad range of linear and nonlinear systems, and derive the associated control laws. The MRAC-RL framework utilizes an inner-loop adaptive controller that allows a simulation-trained outer-loop policy to adapt and operate effectively in a test environment, even when parametric model uncertainty exists. We demonstrate that the MRAC-RL approach improves upon state-of-the-art RL algorithms in developing control policies that can be applied to systems with modeling errors.