In recent times, reinforcement learning has produced baffling results when it comes to performing control tasks with highly non-linear systems. The impressive results always outweigh the potential vulnerabilities or uncertainties associated with the agents when deployed in the real-world. While the performance is remarkable compared to the classical control algorithms, the reinforcement learning-based methods suffer from two flaws, robustness and interpretability, which are vital for contemporary real-world applications. The paper attempts to alleviate such problems with reinforcement learning and proposes the concept of model-assisted reinforcement learning to induce a notion of conservativeness in the agents. The control task considered for the experiment involves navigating a CrazyFlie quadrotor. The paper also describes a way of reformulating the task to have the flexibility of tuning the level of conservativeness via multi-objective reinforcement learning. The results include a comparison of the vanilla reinforcement learning approaches and the proposed approach. The metrics are evaluated by systematically injecting disturbances to classify the inherent robustness and conservativeness of the agents. More concrete arguments are made by computing and comparing the backward reachability tubes of the RL policies by solving the Hamilton-Jacobi-Bellman partial differential equation (HJ PDE).

翻译：暂无翻译