Enhancing AutoMPC for efficient offline optimal control of robots through state constraints, improved system identification models and trajectory tracking controllers

Jeong, Dohun

Permalink

https://hdl.handle.net/2142/120584 Copy

Description

Title

Enhancing AutoMPC for efficient offline optimal control of robots through state constraints, improved system identification models and trajectory tracking controllers

Author(s)

Jeong, Dohun

Issue Date

2023-05-03

Director of Research (if dissertation) or Advisor (if thesis)

Hauser, Kris K

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Robotics
• Machine Learning
• Control

Abstract

Bringing up a robot can be a long, expensive, and dangerous process. This process becomes even more challenging when we build a controller using ma- chine learning techniques. Challenges include: 1) safely collecting a large and diverse enough dataset to the scale that expressive functional approxi- mators such as deep neural networks can capture the complex and nonlinear dynamics of the robot while trading off accuracy with computational load, 2) designing a controller that can reason about both short term constraints and long-term behavior to create an optimal strategy, 3) tuning the parameters of the policy to exhibit desirable behavior when deployed on a physical robot. While areas like reinforcement learning and optimal control address some of these problems, AutoMPC can address all of the above and create a con- troller without any physical interaction beyond the collection of a dataset. In this thesis, several additional features are introduced to make the AutoMPC library more applicable to a variety of robot domains. These include 1) the ability to bound observations in iLQR to ensure safe operation of the robot and prevent hallucination in simulations, 2) a reference trajectory tracking controller, and 3) a modified class of system identification model derived from multilayer perceptrons that uses history of states to predict the robot’s next state. These features are tested on a variety of tasks, including OpenAI gym tasks such as HalfCheetah and CartPole system, along with physical robot tasks on an underwater soft robot arm. This is the first successful application of AutoMPC pipeline to a physical robot, and it outperforms re- cent learning-based methods for creating an optimal controller offline. These features will be included in the public release of the 0.2 version release of AutoMPC, to bring an efficient and scalable solution to data-driven control to the wider research community.

Graduation Semester

2023-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/120584

Copyright and License Information

Copyright 2023 Dohun Jeong

Owning Collections