[2E6-GS-8] Robot and real worlds:

[2E6-GS-8-02] Learning-based Optimal Control Using Generative Models to Compensate for Model Prediction Errors

〇Taichi Hirano¹, Rin Takano¹ (1. NEC Corporation)

Keywords:Predictive models, Optimal Control, Uncertainty

In optimal control using learned transition models, discrepancies may arise between the predicted state and actual system state transitions. This could potentially lead to degraded control performance when handling control inputs and state transitions in out-of-distribution scenarios. In this study, we propose a method that prevents control input and state transitions from deviating from the learning data distribution by leveraging generative models within sampling-based optimal control techniques. This approach allows for the optimization of control inputs while avoiding regions with significant modeling errors, and is expected to result in improved control performance. The effectiveness of the proposed method is demonstrated through simulations.