Optimization for contact-rich robotic manipulation with high-fidelity geometry

Zhang, Mengchao

Permalink

https://hdl.handle.net/2142/124308 Copy

Description

Title

Optimization for contact-rich robotic manipulation with high-fidelity geometry

Author(s)

Zhang, Mengchao

Issue Date

2024-04-24

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

Hauser, Kris Karl

Doctoral Committee Chair(s)

Dullerud, Geir Eirik

Committee Member(s)

• Kim, Joohyung
• Johnson, Aaron M.

Department of Study

Mechanical Sci & Engineering

Discipline

Mechanical Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Robotic Manipulation
• Optimization
• Contact-rich

Abstract

Contact interactions are intricate and pervasive phenomena encountered in various real-world scenarios. Humans and other organisms naturally navigate and interact with their surroundings through physical contact, in contrast to contemporary robots that avoid touching things as much as possible and shy away from complex manipulations such as pushing, pivoting, sliding, and rolling objects. In bridging the gap between human-like contact interactions and current robotic capabilities, this thesis makes foundational contributions through introducing a novel computational model that employs a semi-infinite/infinite programming approach to effectively address the intricate nature of pervasive contact scenarios. The proposed model demonstrates applicability in various contexts, including object's 6D pose estimation within cluttered environments, contact-rich stable grasp pose optimization, and contact-rich manipulation trajectory optimization involving complex-shaped objects. In each of these scenarios, we address the problem that contact is an infinite phenomena involving continuous regions of interaction, which requires a discrete approximation to solve. Contrary to previous methods which need to discretize contacting surfaces a priori into a finite set of contact points, the proposed method operates directly on the continuous underlying geometry, and it dynamically identifies a finite set of constraints essential for problem resolution. As a result, the subsequent solving process is rendered feasible and tractable. Additionally, to enable robust execution of planned trajectories, this thesis introduces two innovative techniques that combine the strengths of model-based planning and reinforcement learning. The first technique, Plan-Guided Reinforcement Learning, uses planned trajectories implicitly as demonstrations. It trains a control policy to mimic the demonstrated motion style while accomplishing the specified task. The policy synthesized in this way exhibits robustness in the face of uncertainties stemming from model parameter inaccuracies and the object's configuration. The second technique, Planned-Contact Informed Policy, utilizes the planned trajectory explicitly. The control policy takes features extracted from the planned trajectory as input, enabling it to anticipate changes in contact modes and adjust its actions accordingly. Consequently, control policies trained in this manner is capable of stabilizing the execution of similar trajectories planned for the same task type. All the contributions presented in this thesis constitute fundamental building blocks towards automated robotic manipulation.

Graduation Semester

2024-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2024 Mengchao Zhang

Owning Collections