Model-based RL with High Dimensional Observations using MPPI and Deep Path-cost Predictor

Abstract

In this paper, we propose a model-based reinforcement learning framework combining Model Predictive Path Integral (MPPI) with a Deep Path-cost Predictor that outputs a state-trajectory cost given an image sequence and a control input sequence as input. We validate the effectiveness of the proposed method by carrying out 2DOF robot arm reaching tasks with multiple targets in simulation.