From Bipedal Walking to Quadrupedal Locomotion: Full-Body Dynamics Decomposition for Rapid Gait Generation

Wen-Loong Ma, Aaron D. AmesICRA 2020 · DOI

TL;DR

This paper introduces a novel method to rapidly generate dynamic quadrupedal gaits by decomposing the full-body dynamics into two simpler bipedal systems. This decomposition allows for significantly faster gait generation, enabling real-time planning and more agile quadrupedal robots.

What Problem Does This Paper Address?

Quadrupedal robots are known for their robust and agile locomotion, often surpassing bipedal robots in real-world scenarios. However, generating dynamic gaits for quadrupeds using their full-body, complex dynamics is computationally intensive. Existing methods often rely on simplified models or extensive parameter tuning, which can be cumbersome. On the other hand, bipedal robots, despite having fewer degrees of freedom, have benefited from more rigorous, formal gait generation techniques based on their full-body dynamics, particularly using the Hybrid Zero Dynamics (HZD) framework. This paper aims to bridge this gap by translating the benefits of formal bipedal gait generation to quadrupeds, enabling rapid, full-body dynamics-based gait planning without sacrificing model fidelity.

Methods

The core of this paper is the decomposition of the quadrupedal robot's full-body dynamics into two independent, yet coupled, bipedal systems. This is achieved by treating the quadruped as two constrained 3D bipedal robots. The authors leverage the Hybrid Zero Dynamics (HZD) framework to rapidly generate gaits and controllers for these decomposed bipedal systems. The key steps involve:

  • Dynamics Decomposition: The nonlinear, hybrid dynamics of the quadruped are mathematically split into continuous and discrete dynamics for two bipedal subsystems. This decomposition is carefully formulated to maintain the essential coupling between the bipeds.
  • Bipedal Gait Generation: The HZD framework, combined with efficient collocation methods, is applied to the decomposed bipedal systems. This allows for the rapid generation of stable walking gaits and controllers for each biped.
  • Gait Recomposition: The generated bipedal gaits are then recomposed to create dynamic quadrupedal locomotion. This process involves using the bipedal trajectories to define the quadruped's overall motion.
  • Optimization Framework: A nonlinear programming (NLP) formulation, utilizing direct collocation (specifically Hermite-Simpson), is employed for efficient gait generation. This approach avoids explicit matrix inversions, contributing to the speedup.

Datasets

Not discussed in the paper. The paper focuses on a specific quadrupedal robot, the Vision 60, and uses its dynamics for simulation and experimental validation.

Key Insights and Ideas

  • Decomposition as a Key to Speed: The most significant insight is that decomposing a complex quadrupedal system into simpler bipedal components dramatically reduces computational complexity, enabling gait generation in seconds rather than minutes or hours.
  • Bridging Bipedal Rigor and Quadrupedal Agility: The method successfully transfers the formal guarantees and rapid generation capabilities from bipedal gait generation to quadrupeds, allowing for dynamic and agile locomotion without relying on heuristic approaches or simplified models.
  • Unified Control for Dynamic Locomotion: The generated gaits can be executed with a simple, time-based PD controller, demonstrating that even open-loop controllers can achieve dynamic stability when based on full-body dynamics.
  • Real-world Feasibility: The method's success in both simulation and outdoor experiments highlights its practical applicability for real-world quadrupedal robots.

Key Results

  • Rapid Gait Generation: The method can generate quadrupedal gaits (stepping-in-place and ambling) in an average of 3.96 seconds. This is an order of magnitude faster than previous full-model based optimization methods.
  • Dynamic Locomotion Achieved: The generated gaits were successfully implemented on the Vision 60 quadruped in both simulation (MuJoCo) and outdoor experiments, demonstrating stable stepping and ambling at 0.35 m/s.
  • Robustness: The quadruped exhibited robustness against rough terrain, wet surfaces, and surface roots, even with a simple time-based controller.
  • Efficiency: The average time per iteration of the optimization was a mere 0.039 seconds.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the time-based control law provides dynamic stability, it doesn't inherently offer robustness against uncertain terrain dynamics, which is a common challenge for open-loop controllers.
  • The current implementation relies on uploading pre-computed gait parameters, suggesting that real-time, on-board gait generation is a future goal.
  • Future Work:
    • The authors envision using this method for real-time, on-board gait generation, enabling dynamic locomotion behaviors to be realized more readily.
  • Exploring code optimization for on-board systems to achieve true real-time gait generation is a key direction.
  • The ultimate goal is to use this approach to realize a wider variety of dynamic locomotion behaviors on quadrupedal robots.

Summary updated

BibTeX
@inproceedings{ma2020from,
  author = {Wen-Loong Ma and Aaron D. Ames},
  title = {{From Bipedal Walking to Quadrupedal Locomotion: Full-Body Dynamics Decomposition for Rapid Gait Generation}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2020},
  doi = {10.1109/icra40945.2020.9196841},
}

Loading related content...

Figure