Towards adaptative wheeled humanoid control: Leveraging fast parameter estimation and sim-to-real adaptation

Baek, Donghoon

Permalink

https://hdl.handle.net/2142/130123 Copy

Description

Title

Towards adaptative wheeled humanoid control: Leveraging fast parameter estimation and sim-to-real adaptation

Author(s)

Baek, Donghoon

Issue Date

2025-06-10

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

Ramos, Joao

Doctoral Committee Chair(s)

Dullerud, Geir

Committee Member(s)

• Kim, Joohyung
• Yim, Justin
• Li, Yunzhu

Department of Study

Mechanical Sci & Engineering

Discipline

Mechanical Engineering

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

Humanoid, Teleoperation, Parameter Estimation, Sim-to-Real

Abstract

Despite promising advancements in robotics, robots are still far from achieving humanlevel capabilities, particularly when it comes to interacting with their environments. In contrast, humans are surprisingly good at adapting to the environment. On of the key secrets behind this capability comes from that we are already familiar with our surrounding environment and could use our intuition based on this information. This raises the question: ”What if we could quickly provide environmental information to a robot?” Furthermore, how beneficial would it be to give this information to robots in a more explicit manner? With those questions in mind, in this work, we propose ‘fast physical parameter estimation’ and ‘model and learning-based control framework’ to explicitly utilize the estimated physical parameters. Specifically, we focus on building adaptive wheeled humanoid control targeting our customized wheeled humanoid robot, SATYRR. Our framework comprises three key components: 1) sim-to-real adaptation, 2) sampling or learning-based parameter estimation, and 3) model or learning-based control. Sim-to-real adaptation enables accurate data and sample collection without needing to operate the physical hardware for specific tasks. This also provides the capability of decoupling the robot dynamics and object dynamics. Sampling and learning-based estimators are typically faster than optimization methods, relying solely on the robot’s proprioception without requiring additional sensors. Finally, the model or learning-based controllers—whether reinforcement learning, MPC, or other optimization-based approaches—can explicitly integrate the estimated parameters into their control formulations. We demonstrated our framework in physical hardware and showed its benefit in enhancing the control performance while making the system more adaptive. The key findings of this work are: 1) Explicitly incorporating environmental physical parameter estimation into the controller enhances adaptability and accuracy. 2) Even if estimation is not highly accurate, a fast estimation process still benefits the controller. 3) From a sim-to-real perspective, critical factors vary by the desired accuracy level: controller gain and cycle are crucial for position/velocity accuracy, while model structure and parameter accuracy are vital for torque-level precision. 4) In teleoperation, applying estimation as haptic feedback or shared autonomy improves task performance and reduces the operator’s workload, especially in complex, fast-paced scenarios.

Graduation Semester

2025-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/130123

Copyright and License Information

Copyright 2025 Donghoon Baek

Owning Collections