Robust Footstep Planning and LQR Control for Dynamic Quadrupedal Locomotion

Guiyang Xin, Songyan Xin, Oguzhan Cebe, Mathew Jose Pollayil, Franco Angelini, Manolo Garabini, Sethu Vijayakumar, Michael N. MistryIEEE RA-L 2021 · DOI

TL;DR

This paper presents a robust framework for dynamic quadrupedal locomotion by combining a high-frequency model predictive control (MPC) for footstep planning with an LQR controller for base motion. This approach enables faster, more responsive locomotion and improved balance, even in the face of disturbances and on varied terrain.

What Problem Does This Paper Address?

Legged robots, particularly quadrupeds, are increasingly being deployed in real-world scenarios. However, achieving robust and dynamic locomotion remains a significant challenge. Existing methods often decouple motion planning and control, running at different frequencies, which can limit responsiveness to unexpected events. Furthermore, controlling underactuated systems like quadrupedal robots, especially during dynamic gaits like trotting, requires sophisticated control strategies that can handle system dynamics and satisfy constraints (e.g., friction cones, torque limits). This paper aims to bridge this gap by proposing a unified, high-frequency framework that integrates planning and control for improved robustness and dynamic capabilities.

Methods

The core of the proposed framework is a two-pronged approach:

  • High-Frequency Footstep Planning (MPC): The paper utilizes a Model Predictive Control (MPC) approach for footstep planning. This MPC operates at a high frequency (400 Hz) and is formulated as a Quadratic Program (QP) that minimizes a cost function related to the robot's Center of Mass (CoM) trajectory and Zero Moment Point (ZMP) positions. Crucially, it uses a simplified Linear Inverted Pendulum Model (LIPM) to predict future CoM states, allowing for fast computation. This high update rate enables the robot to quickly replan its footsteps in response to disturbances or changes in desired motion.
  • LQR Control for Base Motion: For controlling the robot's base (body), the paper employs a Linear Quadratic Regulator (LQR) controller. This LQR controller is designed to track desired base trajectories in a "constraint-free" subspace, while physical constraints (like torque limits and friction cones) are handled in an orthogonal "constraint" subspace using projected inverse dynamics. The LQR controller is derived from a linearized model of the robot's dynamics and is designed to exploit dynamic coupling effects, which is particularly beneficial for underactuated systems. The LQR gains are updated at the same high frequency as the MPC.

The framework also incorporates:

  • Projected Inverse Dynamics: This technique allows for separating the control problem into constraint-free and constraint subspaces, enabling the use of classical controllers like LQR while still satisfying hard physical constraints.
  • Swing Leg Impedance Control: While the paper focuses on the LQR for base control, it mentions that swing legs are controlled using impedance controllers, as in previous work.

Datasets

Not discussed in the paper. The experiments were conducted on a real quadruped robot (ANYmal) and in simulation.

Key Insights and Ideas

  • High-frequency MPC is key for robustness: The paper strongly emphasizes that running the footstep planner at a high frequency (400 Hz) is critical for achieving robustness against disturbances. This allows for rapid replanning of footholds, enabling the robot to recover quickly from unexpected pushes or changes in motion.
  • LQR's strength in underactuated systems: The use of LQR for base control, particularly its ability to exploit dynamic coupling through non-diagonal gain matrices, is highlighted as a significant advantage for controlling underactuated robots like quadrupeds. This contrasts with simpler controllers that might rely on hand-tuned diagonal gains.
  • Separation of concerns with projected inverse dynamics: The framework's ability to handle physical constraints in a separate subspace from the primary trajectory tracking (via LQR) simplifies the controller design and allows for robust performance without complex constraint handling within the LQR itself.
  • Versatility with different foot types: The approach is demonstrated to be effective with both standard spherical feet and adaptive feet, suggesting its generalizability to different hardware configurations.

Key Results

  • Fast Footstep Planning: The MPC planner can be solved within a control cycle of 2.5 ms, enabling high-frequency updates.
  • Improved Balance: The LQR controller allowed for a longer swing phase duration (0.63s) compared to the default ANYmal controller (0.42s) and a previous impedance controller (0.4s), indicating improved balance control.
  • Disturbance Rejection: Experiments showed that the robot could recover from significant disturbances (e.g., being kicked) within one or two steps, with minimal changes in orientation.
  • Terrain Traversability: The system demonstrated robustness on rough outdoor terrain with adaptive feet, suggesting the framework can handle environmental uncertainties.
  • Speed Limitations: While simulation achieved speeds of 1.2 m/s, the real robot's maximum trotting speed was 0.6 m/s, attributed to model errors and hardware limitations.

Limitations and Future Work

  • Height Changing: The current approach does not explicitly address changing the robot's height, which would be necessary for tasks like climbing stairs.
  • Model Errors: While the framework is robust to some model errors, significant discrepancies can still lead to poor trajectory tracking, especially on very slippery terrain.
  • Future Work: The authors suggest adapting the planner to incorporate terrain information for handling slopes and stairs. They also propose using adaptive feet to measure local ground inclination for improved terrain awareness.

Summary updated

BibTeX
@article{xin2021robust,
  author = {Guiyang Xin and Songyan Xin and Oguzhan Cebe and Mathew Jose Pollayil and Franco Angelini and Manolo Garabini and Sethu Vijayakumar and Michael N. Mistry},
  title = {{Robust Footstep Planning and LQR Control for Dynamic Quadrupedal Locomotion}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2021},
  doi = {10.1109/lra.2021.3068695},
}
Figure