Hierarchical Reduced-Order Model Predictive Control for Robust Locomotion on Humanoid Robots

Adrian B. Ghansah, Sergio A. Esteban, Aaron D. AmesHumanoids 2025 · DOI

TL;DR

This paper introduces a novel hierarchical control framework for humanoid robots that combines a high-level, computationally efficient planner using reduced-order models with a lower-level, more detailed model predictive controller. This approach enables robust locomotion across diverse terrains and improved disturbance rejection by intelligently integrating arm and torso dynamics.

What Problem Does This Paper Address?

Deploying humanoid robots in real-world environments presents a significant control challenge due to their complex, high-degree-of-freedom, and underactuated dynamics. Traditional methods often struggle with computational expense, making real-time control difficult, especially when interacting with unpredictable environments. While reduced-order models (ROMs) offer computational efficiency, they typically simplify the robot's dynamics, neglecting important aspects like arm and torso movements that contribute to stability and robustness. This paper aims to bridge this gap by developing a hierarchical control strategy that leverages the speed of ROMs for planning while incorporating crucial whole-body dynamics for enhanced performance.

Methods

The core of the proposed solution is a two-level hierarchical control architecture:

  • High-Level Planner (Nonlinear MPC on ALIP): This layer uses a nonlinear Model Predictive Control (NMPC) approach on the Angular Linear Inverted Pendulum (ALIP) model's step-to-step (S2S) dynamics. This allows for efficient optimization of step lengths, step periods, and ankle torques at a relatively low frequency (40 Hz). The ALIP model is a simplified representation that captures the essential dynamics of bipedal locomotion.
  • Mid-Level Controller (Linear MPC on Decomposed SRB): The optimized step plan from the high-level planner is translated into reference trajectories for a linear Model Predictive Control (MPC) framework. This framework utilizes a Single Rigid Body (SRB) model, but crucially, it's decomposed to include simplified arm and torso dynamics. This decomposition allows the controller to generate forces and moments from the arms and torso to actively stabilize the robot, particularly against yaw disturbances, without sacrificing the linearity required for efficient MPC. The mid-level controller operates at a higher frequency (500 Hz).

Key novelties include:

  • Joint optimization of step lengths and periods in the high-level NMPC.
  • A decomposed SRB model that incorporates arm and torso dynamics while maintaining linear constraints for MPC.
  • A hierarchical structure that separates complex, non-convex optimization (step planning) from fast, linear optimization (whole-body control).

Datasets

Not discussed in the paper. The evaluation is primarily based on simulation and hardware experiments on a Unitree G1 humanoid robot.

Key Insights and Ideas

  • Hierarchical Control is Key: Separating the problem into high-level planning (what steps to take) and mid-level control (how to execute those steps while considering full dynamics) is crucial for achieving both efficiency and robustness.
  • Arm and Torso Dynamics Matter: Incorporating even simplified arm and torso dynamics into the mid-level controller significantly improves the robot's ability to reject external yaw disturbances, leading to more stable locomotion.
  • Adaptive Step Timing Boosts Robustness: Allowing the NMPC to adapt step periods and lengths dynamically, rather than using fixed values, dramatically increases success rates in push recovery scenarios. This suggests that flexibility in stepping is vital for handling unexpected perturbations.
  • ROMs as a Foundation: Reduced-order models, like the ALIP, provide a computationally feasible basis for high-level planning, and their outputs can be effectively translated into references for more complex, whole-body controllers.

Key Results

  • Improved Push Recovery: Adaptive step timing increased successful push recovery rates by 36% compared to fixed step periods.
  • Enhanced Disturbance Rejection: The proposed DSRB-MPC (with upper body control) consistently reduced yaw deviations more effectively than a standard SRB-MPC when subjected to torso yaw disturbances.
  • Real-time Performance: The high-level planner runs at 40 Hz, and the mid-level MPC and low-level command mapper run at 500 Hz on the robot's onboard mini-PC, demonstrating practical feasibility.
  • Robust Locomotion: The framework demonstrated robust locomotion across diverse terrains, including grass, stone pavement, and uneven gym mats, in hardware experiments.

Limitations and Future Work

  • ALIP Model Simplification: The authors acknowledge that the high-level ALIP model does not fully capture the richer torso and arm dynamics that are represented in the DSRB model. This disconnect could limit the planner's ability to fully exploit the capabilities of the lower-level controller.
  • Future Directions: The paper suggests exploring closer integration between the high-level planner and the DSRB model. Additionally, extensions to more dynamic gaits like running and hopping are proposed as future work.

Summary updated

BibTeX
@inproceedings{ghansah2025hierarchical,
  author = {Adrian B. Ghansah and Sergio A. Esteban and Aaron D. Ames},
  title = {{Hierarchical Reduced-Order Model Predictive Control for Robust Locomotion on Humanoid Robots}},
  booktitle = {Humanoids},
  year = {2025},
  doi = {10.1109/humanoids65713.2025.11203156},
}
Figure