Robot Control Using Path Integral Policy Improvement and Deep Dynamics Models

Abstract: Robotics is an interdisciplinary field that integrates computer science, electrical engineering, mechanical engineering, control engineering and other related fields. As the quick development of these fields, people have been building more complex robots with more advanced control strategies in order to solve more challenging tasks. In addition, it is always a target for researchers to achieve autonomous operation of robots so that the manpower can be saved and the robot can work in harsh environment like on Mars. In this project, I focus on the trajectory planning problem of a unicycle model running in 2D environment. I choose Path Integral Policy Improvement (PI2) control algorithm in this project as the main study object. And Model Predictive Control (MPC) is chosen as a reference in order to be compared with PI2 to evaluate the performance of PI2. In order to simulate the tasks that the robot needs to handle in practice, I use obstacles to represent the complex environment and I use Signal Temporal Logic (STL) to represent the complex tasks. Furthermore, I also incorporate the deep dynamics model in the project so that the the method put forward in this project is able to handle complex robot models and complex working environments. To evaluate the performances of PI2 and MPC, five criteria are put forward in this project. Finally, based on the evaluation results, possible improvement and future research are proposed.