Uncertainty-Aware Robot Control via Model-Based Reinforcement Learning
Zhibo Zhou · 2024
In recent years, model-based reinforcement learning methods have been extensively studied in the field of robot control, due to their high sampling efficiency. However, current model-based methods often don’t take into account the performance drop caused by uncertainty in robot control scenarios. To address this issue, we propose Uncertainty-Aware Model Learning (UAML). We quantify the uncertainty of the model based on model ensembles and use it to assess the accuracy of generated data. Then, we introduce a sampling-weighted mechanism to utilize the generated data. We validate the effectiveness of our approach through four robot control tasks on a high-precision simulation platform. Our method consistently outperforms state- of-the-art model-based methods.