Dynamics Consensus between Centroidal and Whole-Body Models for Locomotion of Legged Robots

Rohan Budhiraja, Justin Carpentier, Nicolas MansardICRA 2019 · DOI

TL;DR

This paper introduces a novel mathematical framework using the Alternating Direction Method of Multipliers (ADMM) to enforce dynamic consensus between simplified centroidal robot models and full whole-body dynamics. This approach allows for more robust and coherent locomotion planning by ensuring that the reduced-order and full-order models agree on key dynamic quantities like momentum.

What Problem Does This Paper Address?

Locomotion for legged robots is a complex optimal control problem due to the high dimensionality and non-convexity of the robot's dynamics. A common simplification is to use reduced-order models, such as the centroidal model, which captures the robot's overall momentum. However, ensuring that the solutions derived from these simplified models are dynamically consistent with the full, actuated whole-body dynamics is a significant challenge. Existing methods often rely on approximations or heuristics, lacking a rigorous mathematical foundation for achieving this "dynamics consensus." This paper aims to bridge this gap by providing a principled way to reconcile the centroidal and whole-body dynamics.

Methods

The core of the paper's contribution lies in the application of the Alternating Direction Method of Multipliers (ADMM) to solve the coupled optimal control problem for locomotion. ADMM is a powerful optimization technique for problems that can be decomposed into separable subproblems with linear constraints. The authors first formulate the locomotion problem by splitting it into two hierarchical stages: one for the centroidal dynamics (under-actuated) and another for the whole-body dynamics (fully actuated).

They then define residual functions that capture the coupling constraints between these two subproblems, specifically related to the Center of Mass (CoM) position and Angular Momentum (AM). ADMM is used to iteratively solve these subproblems, with the dual variables acting as a feedback mechanism to enforce consensus. The paper also discusses practical implementation details, including the use of existing solvers for the centroidal and whole-body dynamics, and how to initialize the dual variables for faster convergence.

Datasets

Not discussed in the paper. The paper focuses on a simulation-based validation using a model of the HRP-2 robot.

Key Insights and Ideas

  • ADMM as a principled consensus mechanism: The paper demonstrates that ADMM provides a mathematically rigorous way to enforce agreement between simplified and full robot dynamics, moving beyond heuristic approaches.
  • Feedback through dual variables: The dual variables in ADMM act as an integral controller, collecting residues from constraint violations and driving the subproblems towards a common solution. This is crucial for reconciling the different levels of abstraction.
  • Splitting the problem: The authors show that a clean splitting of the locomotion problem into centroidal and Lagrangian dynamics is possible, and that ADMM can effectively manage the coupling constraints that arise from this split.
  • Importance of AM consistency: The paper highlights that inconsistencies in Angular Momentum (AM) between centroidal and whole-body models are a major hurdle, especially for dynamic movements, and their method directly addresses this.
  • Practical implementation: The paper provides insights into how to integrate ADMM with existing locomotion solvers, making it a practical extension rather than a complete overhaul.

Key Results

The paper validates its approach through simulations of walking on the HRP-2 robot. Key findings include:

  • Rapid convergence: The ADMM-based solver achieves a consensus between centroidal and whole-body trajectories within a small number of iterations (around 3-10 iterations for significant convergence).
  • Matching trajectories: Figures demonstrate that the CoM trajectories from the centroidal and whole-body subproblems closely match after just a few iterations.
  • Decreasing residuals: The total residuals of the matching constraints decrease rapidly, indicating that the ADMM approach effectively enforces the desired dynamic consensus.
  • Feasible walking motion: The method successfully generates a walking motion sequence for the HRP-2 robot.

Limitations and Future Work

  • Non-convexity: While ADMM is theoretically guaranteed for convex problems, its application to non-convex problems like robot locomotion is empirical. The authors acknowledge that it acts as a local optimizer in such cases, but note that linearization can help.
  • Contact planner dependence: The current locomotion pipeline assumes a pre-defined contact sequence from an upstream planner. The authors suggest that ensuring consistency between this contact planner and the whole-body optimizer is a crucial area for future research.
  • Initialization: While the paper discusses a trick to potentially avoid an extra iteration by initializing dual variables, it was not used in the presented experiments and warrants further evaluation.
  • Real-world validation: The results are presented from simulations; extending this to real-world robot experiments would be a natural next step.

Summary updated

BibTeX
@inproceedings{budhiraja2019dynamics,
  author = {Rohan Budhiraja and Justin Carpentier and Nicolas Mansard},
  title = {{Dynamics Consensus between Centroidal and Whole-Body Models for Locomotion of Legged Robots}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2019},
  doi = {10.1109/icra.2019.8793878},
}
Figure