A simulation framework for simultaneous design and control of passivity based walkers

Guilhem Saurel, Justin Carpentier, Nicolas Mansard, Jean-Paul Laumond

TL;DR

This paper presents a high-performance simulation framework that simultaneously optimizes a bipedal robot's physical design (e.g., limb lengths, mass distribution) and its control parameters. By leveraging efficient rigid-body dynamics and multiple-shooting optimization, the tool can generate energy-efficient walking gaits for complex 3D robots in just a few seconds.

What Problem Does This Paper Address?

Designing "passivity-based" walkers—robots that rely on their natural dynamics and gravity to move—is a classic "chicken and egg" problem. The robot's physical shape (its "design") dictates its natural gait, while the controller must be tuned to exploit those specific physics.

Traditionally, researchers design the hardware first and then struggle to find a controller, or vice versa. Furthermore, many existing simulation tools are either too slow for rapid iteration or limited to simple 2D "compass" models. The authors aim to bridge this gap with a generic framework that handles complex 3D architectures and optimizes both design and control in a single, fast computational step.

Methods

The framework treats robot design and control as a single large-scale parametric optimal control problem. Key components include:

  • Dynamics Engine (Pinocchio): The authors use the Pinocchio C++ library, which is optimized for fast forward and inverse dynamics of poly-articulated systems. This allows the framework to automatically compute the complex equations of motion for any kinematic tree without manual derivation.
  • Optimization Solver (MUSCOD-II): They employ a Multiple Shooting method. Unlike "Single Shooting" (which is sensitive to initial guesses) or "Collocation" (which requires very fine grids), Multiple Shooting breaks the trajectory into smaller intervals. This makes it particularly robust for the unstable, hybrid dynamics (continuous swinging mixed with discrete foot-strike impacts) typical of walking.
  • Hybrid Dynamics: The model accounts for continuous motion during the stance phase and instantaneous inelastic impacts when the foot hits the ground.
  • Actuation Modes: The framework can optimize for active control (where joints are powered by motors following cubic polynomials) or passive control (where joints act like springs and dampers).

Datasets

The paper does not use external training datasets. Instead, it relies on:

  • Anthropometric Tables: To ensure the robots remain "human-like," the authors use standard biomechanical data (McConville et al., 1980) to scale mass, center of mass, and inertia based on a reference height and weight.
  • Synthetic Case Studies: The framework is evaluated on six distinct walker models: a basic compass walker, a walker with knees, one with a torso, one with a neck, and a 3D model with arms.

Key Insights and Ideas

  • The "Aha" Moment on Complexity: Counter-intuitively, adding complexity to the robot's body often makes it more efficient. For example, adding knees to a simple compass walker roughly halves the energy cost.
  • Design-Control Coupling: By including physical parameters (like limb length) in the optimization vector alongside control gains, the solver can find "sweet spots" where the robot's body naturally assists the movement, reducing the effort required by the controller.
  • Efficiency of Multiple Shooting: The authors highlight that for multi-phase systems with impacts, multiple shooting is the superior numerical strategy because it keeps the optimization problem low-dimensional while remaining stable.

Key Results

  • Computational Speed: The framework is remarkably fast, converging on optimal designs and gaits in 1.2 to 7.6 seconds on a standard computer.
  • Energy Efficiency (Cost of Transport - CoT):
    • Knees: Adding knees to the base model reduced the CoT from 0.1007 to 0.0544.
  • Torso: Adding a fixed torso further reduced the CoT by 40%.
  • 3D Capability: The framework successfully optimized a 3D walker with actuated arms, demonstrating that it can handle the lateral balance requirements that 2D models ignore.
  • Active vs. Passive: While active controllers achieved lower CoT, the passive (spring-damper) optimization was significantly faster to compute due to having fewer parameters.

Limitations and Future Work

  • Stability Proofs: While the framework finds balanced, periodic gaits, it does not explicitly check for "orbital stability" (the ability to recover from a push). The authors suggest integrating Poincaré map analysis in the future.
  • Contact Modeling: The current model uses "punctual" (point) contact. Real-world walkers often have curved or rolling feet, which the authors plan to implement to improve gait realism.
  • Open Source Availability: The solver used (MUSCOD-II) is closed-source, though the authors note that the open-source ACADO toolkit offers similar functionality.

Summary updated

Abstract

In this paper, we propose a simulation framework which simultaneously computes both the design and the control of bipedal walkers. The problem of computing a design and a control is formulated as a single large-scale parametric optimal control problem on hybrid dynamics with path constraints (e.g. non sliding and non slipping contact constraints). Our framework relies on state-of-the-art numerical optimal control techniques and efficient computation of the multi-body rigid dynamics. It allows to compute both the parametrized model and the control of passive walkers on different scenarios, in only few seconds on a standard computer. The framework is illustrated by several examples which highlight the interest of the approach.
BibTeX
@inproceedings{saurel2016simulation,
  author = {Guilhem Saurel and Justin Carpentier and Nicolas Mansard and Jean-Paul Laumond},
  title = {{A simulation framework for simultaneous design and control of passivity based walkers}},
  booktitle = {SIMPAR},
  year = {2016},
  doi = {10.1109/simpar.2016.7862383},
}

Loading related content...

Figure