Model-Based Reinforcement Learning with LSTM Networks for Non-Prehensile Manipulation Planning

Abstract

Solving non-prehensile manipulation tasks requires domain knowledge involving various interactions such as switching contact dynamics between the robot and the object, and the object-environment interactions. This results in a switched nonlinear dynamic system governing the physical interactions between the object and the environment. In this paper, we propose an interactive learning framework that allows a robot to autonomously learn and model an unknown object's dynamics, as well as utilise the learned model for efficient planning in completing re-positioning tasks using non-prehensile manipulation. First, we model the overall object dynamics using a Long Short-Term Memory (LSTM) neural network. We then assimilate the learned model into the Monte Carlo Tree Search (MCTS) algorithm with a dense reward function to generate an optimal sequence of push actions for task completion. We demonstrate the framework in both simulated and real robot that pushes objects on a table.