11/01/2023
By Danielle Fretwell

The Francis College of Engineering, Department of Mechanical Engineering, invites you to attend a Doctoral Dissertation Proposal defense by Yuan Gao on: Time-Varying Control and State Estimation for Bipedal Locomotion.

Candidate Name: Yuan Gao
Degree: Doctoral
Defense Date: Nov. 15, 2023
Time: 9 to 11 a.m.
Location: Ball Hall 323

Committee:

  • Advisor: Christopher Niezrecki, Distinguished University Professor, Mechanical Engineering, UML
  • Co-Advisor: Yan Gu, Associate Professor, School of Mechanical Engineering, Purdue University
  • Kshitij Jerath, Assistant Professor, Mechanical Engineering, UML
  • Kelilah Louise Wolkowicz, Assistant Professor, Mechanical Engineering, UML
  • Xingye Da, General Manager, XPeng Robotics

Brief Abstract:

This thesis focuses on advancing bipedal robot control and state estimation within the context of dynamic environments. It encompasses several key stages, starting with the establishment of Global-Position Tracking (GPT) control for multi-domain walking, where walking involves alterations in contact conditions or actuator configurations during motion. This initial phase lays the foundation for precise trajectory tracking and stability, essential for navigating rapid changing environments. Developing a tracking controller for bipedal robots poses unique challenges due to the intricate, time-varying, and hybrid nature of robot dynamics, particularly in the context of multi-domain walking, encompassing various phases with distinct actuation characteristics. In response to this challenge, a continuous-phase GPT control law for multi-domain walking is introduced, which provides provable exponential convergence of the entire error state within full and over actuation domains, as well as the directly regulated error state within the underactuation domain. The simulation results underscore the accuracy and convergence rate of the proposed control approach.

Subsequently, the thesis delves challenges of locomotion on Dynamic Rigid Surfaces (DRSes), mirroring real-world settings like trains and ships. Specific control and state estimation methods are introduced for bipedal robots navigating these surfaces.

The thesis explores the extension of state estimation techniques from static terrains to DRSes. This extension aims to enhance the accuracy of state estimation with DRS motion involved, leading to more informed decision-making for improved control performance. An innovative Invariant Extended Kalman Filter is introduced for estimating the robot's pose and velocity during DRS locomotion using common sensors found on legged robots, such as inertial measurement units (IMU), joint encoders, and RGB-D cameras. The filter explicitly accounts for the nonstationary surface-foot contact point and the hybrid robot behaviors. It also exhibits attractive properties like guaranteed convergence in the absence of IMU biases. Furthermore, the observability of state variables is analyzed, revealing the impact of DRS movement on state observability. Experimental validations featuring a Digit humanoid robot navigating a pitching treadmill affirm the efficacy of the proposed filter under conditions including sensor noise, biases, estimation errors, and DRS movement.

The last accomplishment of this research involves the introduction and evaluation of novel angular momentum-based linear inverted pendulum (ALIP) based control strategies. These strategies are meticulously crafted to ensure stable bipedal walking on DRSes. This final component encompasses a real-time planning and control framework designed specifically for underactuated walking on DRSes featuring known periodic motion. The framework comprises three layers: foot-placement planning, full-body reference generation, and feedback control. By incorporating ALIP model and a DRS forcing input, the framework stabilizes the hybrid, linear, time-varying ALIP model, ensuring stable walking. The walking pattern generator produces smooth foot placement transitions using Bézier polynomials encoded by a time-based phase variable, while the lower layer implements feedback controllers to drive the robot to its desired state. Validations using the Digit robot, both in simulations and hardware experiments, support the framework's applicability and potential for complex system control on DRSes.

In summary, this thesis progresses through the stages of global-position tracking control, adaptation to DRSes through advanced state estimation techniques and the development of LIP-based control strategies. These endeavors collectively empower bipedal robots to navigate dynamic environments with enhanced stability, accuracy, and adaptability, significantly expanding their potential applications.