Motion Decoupling and Composition via Reduced Order Model optimization for Dynamic Humanoid Walking with CLF-QP based Active Force Control

Xiaobin Xiong, Aaron D. AmesIROS 2019 · DOI

TL;DR

This paper presents a novel framework for generating dynamic humanoid walking by decoupling complex 3D motion into simpler planar components. By composing optimized motions from reduced-order models (ROMs) and embedding them onto a full-dimensional robot using Control Lyapunov Function-based Quadratic Programs (CLF-QPs), the authors achieve versatile and stable walking behaviors, including dynamic transitions.

What Problem Does This Paper Address?

Generating dynamic and versatile walking gaits for humanoid robots is a significant challenge. Traditional methods often rely on simplified models like the Linear Inverted Pendulum (LIP) for planning, which can lead to quasi-dynamic motions. While more dynamic models like the Spring-Loaded Inverted Pendulum (SLIP) capture more realistic dynamics, their inherent complexity makes online optimization and control difficult. Furthermore, achieving smooth transitions between different walking behaviors (e.g., from standing to periodic walking) and ensuring robust force control for stable ground interaction remain open problems. This paper aims to bridge this gap by leveraging the strengths of both simplified and complex ROMs to generate dynamic walking with robust force control.

Methods

The core of this work lies in a hierarchical approach:

  • Motion Decoupling: The 3D walking motion is first decoupled into periodic and transitional components. Each of these is further broken down into sagittal (forward/backward) and lateral (side-to-side) planar motions.
  • Reduced-Order Models (ROMs):
    • Sagittal Plane (Periodic): An actuated Spring-Loaded Inverted Pendulum (aSLIP) model is used to generate dynamic periodic walking motions.
  • Lateral Plane (Periodic): A Hybrid passive Linear Inverted Pendulum (H-LIP) model is employed for periodic lateral motion.
  • Transitional Motion: Footed versions of the aSLIP and H-LIP models are used for generating fast and smooth transitions between different walking states. These models incorporate foot actuation to better capture dynamic maneuvers.
  • Motion Composition: The optimized planar COM dynamics from the ROMs are composed to create a desired 3D COM trajectory.
  • Dynamics Embedding: A Control Lyapunov Function-based Quadratic Program (CLF-QP) framework is used to embed the desired COM dynamics onto the full-dimensional humanoid robot. This approach directly generates control torques for the robot.
  • Active Force Control: A key innovation is the integration of active force control within the CLF-QP framework. This allows the robot's ground reaction forces to match those generated by the aSLIP model, ensuring smooth transitions between support phases (e.g., from double support to single support) by allowing the normal force to naturally approach zero.

Datasets

Not discussed in the paper. The work focuses on simulation and theoretical development, with a mention of implementing on a lower-limb exoskeleton.

Key Insights and Ideas

  • Decoupling for Complexity Management: Breaking down complex 3D walking into simpler planar ROMs significantly simplifies the optimization and control problem, making dynamic motions more tractable.
  • Synergistic ROM Usage: Combining the strengths of aSLIP (dynamic, natural oscillations) for sagittal motion and H-LIP (linear, easier optimization) for lateral motion provides a balanced approach.
  • Footed ROMs for Transitions: Using foot-actuated ROMs for transitions is crucial for capturing the dynamic maneuvers required for fast and smooth state changes, which point-footed models struggle with.
  • CLF-QPs for Robust Embedding: CLF-QPs provide a powerful framework for ensuring stability and tracking of desired dynamics while also accommodating physical constraints.
  • Force Control for Smooth Transitions: Explicitly matching ground reaction forces to ROM-generated profiles is a critical insight for achieving seamless transitions between different contact phases, preventing abrupt changes in dynamics. The relaxation mechanism for force matching is also important for practical implementation.

Key Results

The paper demonstrates the effectiveness of the proposed framework through simulations on a lower-limb exoskeleton model. Key results include:

  • Successful generation of periodic walking motions using the composed ROM dynamics.
  • Generation of dynamic transitional motions from standing to periodic walking.
  • The CLF-QP with active force control successfully embeds the desired COM dynamics and ground reaction forces onto the robot.
  • Simulations show that the ground normal forces remain within the relaxed bounds, indicating smooth transitions. The impact of foot-ground contact is present but does not destabilize the system due to the force control.

Limitations and Future Work

  • Simulation-Based: The primary results are from simulation. Hardware implementation is planned but not yet presented.
  • Pre-specified Contact Sequence: The transition motion is pre-specified in terms of contact sequence and number of domains. More advanced methods like contact implicit optimization are mentioned as future work.
  • Relaxation in Force Control: While necessary for stability, the relaxation in force matching means the robot's ground reaction forces don't exactly match the ROMs, but rather stay within a defined range.
  • Future Work:
    • Hardware implementation on an exoskeleton.
  • Online gait generation using the ROMs.
  • Realization of CLF-QP with contact force control on hardware.
  • Exploring more advanced contact handling and optimization techniques.

Summary updated

BibTeX
@inproceedings{xiong2019motion,
  author = {Xiaobin Xiong and Aaron D. Ames},
  title = {{Motion Decoupling and Composition via Reduced Order Model optimization for Dynamic Humanoid Walking with CLF-QP based Active Force Control}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2019},
  doi = {10.1109/iros40897.2019.8968215},
}
Figure