An Optimization-Based Locomotion Controller for Quadruped Robots Leveraging Cartesian Impedance Control

Guiyang Xin, Wouter Wolfslag, Hsiu-Chin Lin, Carlo Tiseo, Michael N. MistryFrontiers Robotics AI 2020 · DOI

TL;DR

This paper introduces a novel locomotion controller for quadruped robots that combines analytical Cartesian impedance control with Quadratic Programming (QP). This approach allows robots to achieve precise movements while maintaining compliance and satisfying physical constraints, even in challenging terrains, by intelligently trading off trajectory tracking for stability when necessary.

What Problem Does This Paper Address?

Quadruped robots are increasingly important for tasks in complex environments, but their locomotion is challenged by the need for both precise control and inherent compliance. This compliance is crucial for handling unexpected forces from rough terrain or human interaction. Existing control methods often struggle to balance these competing demands. For instance, some methods require extensive manual tuning of parameters, while others might not effectively handle multiple constraints simultaneously, leading to potential instability like slipping. The paper aims to bridge this gap by providing a unified control framework that analytically derives desired motions and then uses optimization to ensure these motions are physically feasible, even under difficult conditions.

Methods

The core of the proposed controller lies in a two-stage process:

  1. Analytical Cartesian Impedance Control: This stage uses "projected inverse dynamics" to derive an analytical control law. This allows the robot to behave like a mass-spring-damper system in its operational space (e.g., torso position and orientation), enabling it to respond predictably to disturbances. A key advantage is that this method can also estimate external forces and model errors, which can then be used for compensation.
  2. Quadratic Programming (QP) Optimization: The desired torques from the impedance controller are then fed into a QP solver. This QP ensures that the robot's actions satisfy critical physical constraints, such as:
  3. Unilateral constraints: Preventing feet from losing contact with the ground.
  4. Friction cone constraints: Ensuring feet don't slip.
  5. Torque limits: Respecting the physical capabilities of the robot's actuators.

    Crucially, if the desired motion from the impedance controller would violate these constraints, the QP intelligently performs a trade-off. It sacrifices some trajectory tracking accuracy to ensure physical feasibility, preventing slips or other failures. This is a significant improvement over methods that might fail entirely if constraints are violated. The paper also highlights that this approach requires solving only one QP, making it computationally more efficient than hierarchical QP (HQP) methods that often require multiple QPs.

Datasets

Not discussed in the paper. The experiments were conducted on the ANYmal robot.

Key Insights and Ideas

  • Analytical Impedance Control for Compliance and Estimation: Deriving the impedance control law analytically provides a principled way to achieve desired compliance and, importantly, to estimate external forces and model errors. This is a powerful tool for adapting to unknown environments or payloads.
  • QP for Feasible Trade-offs: The QP formulation is key to ensuring physical feasibility. The ability to intelligently trade off trajectory tracking for constraint satisfaction is critical for robust locomotion, especially in uncertain or challenging conditions. This prevents catastrophic failures like slipping.
  • Single QP Efficiency: By integrating the impedance control output directly into a single QP, the controller achieves computational efficiency compared to multi-stage HQP approaches, making it more suitable for real-time applications.
  • Model Error Compensation: The ability to estimate and compensate for model errors (e.g., due to carrying an unknown load) is a significant practical advantage, allowing the robot to maintain stable locomotion even when its dynamics change.

Key Results

  • Computational Efficiency: The proposed controller is computationally faster than HQP controllers, requiring less time per control cycle. For static walking, the proposed controller's total computation time was around 0.24-0.27 ms, while HQP took around 0.43-0.45 ms.
  • External Disturbance Estimation: The controller successfully estimated external pushing forces (around 35N estimated vs. 30N measured), demonstrating its ability to perceive and react to disturbances. HQP controllers, in contrast, struggled to provide accurate disturbance estimates without extensive tuning.
  • Model Error Compensation: When carrying a 10 kg load, the controller significantly reduced position errors after enabling model error compensation, demonstrating its effectiveness in adapting to payload changes.
  • Challenging Terrain Performance: The robot successfully walked on low-friction surfaces (PTFE sheets) and climbed a 30° slope, showcasing the controller's robustness.
  • Trade-off Demonstration: Experiments showed that when walking on low-friction surfaces, the controller intentionally sacrificed some trajectory tracking accuracy (indicated by increased errors in z-direction and deviation between desired and actual torques) to maintain friction cone constraints, preventing slips.

Limitations and Future Work

  • Static Gait Planning: The experiments used a default static gait planner from ANYbotics. Future work could focus on optimizing step planning for dynamic gaits and challenging terrains.
  • Real-time Model Adaptation: While model error compensation was demonstrated, continuous real-time adaptation of the robot's dynamic model is identified as an important upgrade for real-world deployment.
  • Friction Cone Approximation: The paper mentions approximating the friction cone with a linearized pyramid for computational efficiency. While effective, this is an approximation, and more precise handling of the quadratic constraint could be explored.
  • Tuning of Gains: While the analytical approach reduces the need for manual tuning compared to HQP, some gains (like stiffness and damping) still require careful selection, though the paper suggests analytical methods can guide this.

Summary updated

BibTeX
@article{xin2020optimization,
  author = {Guiyang Xin and Wouter Wolfslag and Hsiu-Chin Lin and Carlo Tiseo and Michael N. Mistry},
  title = {{An Optimization-Based Locomotion Controller for Quadruped Robots Leveraging Cartesian Impedance Control}},
  journal = {Frontiers Robotics AI},
  year = {2020},
  doi = {10.3389/frobt.2020.00048},
}
Figure