Robust whole-body motion control of legged robots

Farbod Farshidian, Edo Jelavic, Alexander W. Winkler, Jonas BuchliIROS 2017 · DOI

Introduction

Most whole-body controllers for legged robots rely on the principle of "Inverse Dynamics." In an ideal scenario where the robot's mass and environmental physics are perfectly known, one can calculate the precise torques required at each actuator to achieve a desired movement. However, real-world applications often violate these assumptions. Actuators exhibit internal dynamics and communication delays, ground surfaces can be soft or uneven, and rigid-body models are inherently approximations.

Standard controllers typically assume "ideal" force-tracking actuators. When these robots encounter unpredictable conditions—such as stepping on foam or a moving platform—the mismatch between the model and reality often leads to instability, drift, and eventual failure. This research introduces a robust control architecture designed to handle these uncertainties by moving away from the ideal actuator assumption.

Methods

The proposed architecture achieves robustness through a three-pronged approach: subsystem decomposition, Internal Model Control (IMC), and coequal task decoupling.

Subsystem Decomposition

The robot's complex physics are broken down into three manageable subsystems to simplify control design:

  • CoM Subsystem: Manages the Center of Mass to ensure overall balance.
  • Contact Subsystem: Controls the forces at the feet currently in contact with the ground.
  • Non-contact Subsystem: Manages "swing" limbs moving through the air.
The proposed control architecture featuring subsystem decomposition into CoM, contact, and non-contact subsystems.

Internal Model Control (IMC)

The core innovation for force control is the application of IMC. Rather than a standard feedback loop, IMC utilizes a nominal model of the actuator—accounting for delays and time constants—to predict output. The controller then uses the discrepancy between measured and predicted forces to reject disturbances. This system can be tuned using a single parameter ($\eta_f$) to balance performance against robustness.

Coequal Task Decoupling

While traditional controllers often use strict task hierarchies (e.g., prioritizing balance over leg placement), this method introduces "coequal" decoupling. Tasks are mathematically separated into independent subspaces, ensuring that the contact force controller operates without interference from swing leg movements, preventing "cross-talk" between different robot components.

Experimental Evaluation

The controller was evaluated using both simulation and hardware tests on the HyQ hydraulic quadruped robot. Simulation tests focused on "soft ground" scenarios with low spring constants ($1.0 \times 10^4 N/m$). Hardware experiments included walking on foam-wrapped feet to simulate unpredictable surface stiffness and "blind" balance tests on moving planks and seesaws where the robot had no sensors to detect ground movement.

Key Results

The robust controller demonstrated significant superiority over standard inverse dynamics, particularly in unpredictable environments. In soft-ground simulations, standard inverse dynamics accumulated approximately 10 cm of tracking error, whereas the proposed robust controller maintained an error of under 1 cm.

Comparison of tracking performance between the proposed robust full-body control (left) and standard inverse dynamics (right) on soft ground.

On physical hardware, the HyQ robot successfully navigated foam-covered surfaces. During the moving plank and seesaw experiments, the IMC-based force controller automatically adjusted leg torques to maintain an upright pose despite the lack of ground-movement sensors. The robot maintained its balance with a pitching angle deviation of less than 2 degrees.

Conclusion

By restructuring the controller to expect and handle model errors rather than relying on increasingly complex models, the authors have developed a resilient system for legged locomotion. The integration of IMC for task-space force control provides a systematic method for handling the inherent lag and "sponginess" of real-world interactions. Future work will focus on integrating this architecture with dynamic motion planners to enable reactive footstep adjustments during high-speed maneuvers like trotting.

Summary updated

BibTeX
@inproceedings{farshidian2017robust,
  author = {Farbod Farshidian and Edo Jelavic and Alexander W. Winkler and Jonas Buchli},
  title = {{Robust whole-body motion control of legged robots}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2017},
  doi = {10.1109/iros.2017.8206328},
}
Figure