Torque Saturation in Bipedal Robotic Walking Through Control Lyapunov Function-Based Quadratic Programs

Kevin S. Galloway, Koushil Sreenath, Aaron D. Ames, Jessy W. GrizzleIEEE Access 2015 · DOI

TL;DR

This paper introduces a novel method for bipedal robots to handle torque limitations during walking by integrating these constraints directly into a control Lyapunov function (CLF) framework. By formulating the control problem as a convex optimization, the controller can now respect torque limits without sacrificing stability, leading to smoother and more robust walking.

What Problem Does This Paper Address?

Bipedal robots, like the MABEL robot used in this study, are inherently complex systems with underactuated dynamics. A common approach to achieve stable walking is using Control Lyapunov Functions (CLFs), which provide theoretical guarantees of stability and convergence. However, real-world robots have physical limitations, such as maximum motor torques (torque saturation). Previous CLF-based controllers applied these saturations after calculating the ideal control signal, leading to degraded performance and potential instability when the ideal torque exceeded the physical limits. This paper addresses the critical challenge of proactively incorporating these torque saturation constraints into the CLF control design itself, ensuring that the robot's movements remain stable and predictable even when facing physical limitations.

Methods

The core innovation lies in reformulating the CLF-based control problem as a convex optimization.

  • Control Lyapunov Functions (CLFs): The paper builds upon existing work using CLFs, specifically "rapidly exponentially stabilizing CLFs" (RES-CLFs), to guarantee stability for hybrid systems like bipedal robots.
  • Hybrid Zero Dynamics (HZD): This framework is used to define a lower-dimensional manifold on which stable walking gaits exist.
  • Pointwise Min-Norm Controller: A standard approach to generate a CLF controller that minimizes control effort while satisfying stability conditions.
  • Convex Optimization Formulation: The key novelty is expressing the pointwise min-norm controller as a convex optimization problem. This allows for the direct inclusion of constraints.
  • Handling Saturation:
    • Soft Saturation: Initially, the authors relax the CLF stability condition slightly and introduce "soft" bounds on the control inputs. This allows the controller to trade off perfect adherence to the saturation limit for better CLF performance, penalizing violations.
  • Hard Saturation: Subsequently, they present a formulation that enforces hard saturation bounds, meaning the control output will never exceed the physical limits. This is achieved by ensuring the convex optimization problem can be solved within the strict real-time control loop.

Datasets

Not discussed in the paper. The experiments utilize the MABEL bipedal robot and its associated dynamics model.

Key Insights and Ideas

  • Proactive Saturation Handling is Crucial: Simply clipping control signals after calculation is insufficient for robust CLF-based control. The constraints must be integrated into the optimization problem.
  • Convex Optimization as a Unifying Framework: Framing the CLF control problem as a convex optimization provides a powerful and flexible way to incorporate various constraints, not just torque saturation.
  • Trade-off Between Performance and Saturation: The "soft saturation" approach explicitly models the trade-off between maintaining the CLF's theoretical stability guarantees and respecting physical torque limits.
  • Real-time Feasibility: The use of CVXGEN for code generation is critical, enabling the solution of these complex optimization problems within the tight (1 kHz) real-time control loop required for dynamic walking.

Key Results

  • Stable Walking with Saturation: The proposed method successfully enabled stable walking on the MABEL robot, even with significant torque saturation.
  • Smoother Control: Compared to previous methods, the CLF controller with integrated saturation handling produced smoother control torques.
  • Gradual Performance Degradation: Numerical simulations showed that as saturation bounds became more restrictive, tracking errors increased gradually, indicating a graceful degradation of performance rather than abrupt failure.
  • Experimental Validation: The method was successfully implemented on the MABEL robot, demonstrating its effectiveness in a real-world setting. The hard saturation implementation resulted in torques that were visibly "flattened" at the boundaries.

Limitations and Future Work

  • Limitations:
    • The paper notes that in the hard saturation case, the convex optimization might fail to converge within the strict time constraints, leading to the last computed control signal being held. This is an isolated occurrence and doesn't affect system stability but represents a momentary failure to meet the hard bound.
  • The paper mentions that the bounds can be specified as functions of time or system state, suggesting a path towards dynamic saturation limits (e.g., based on battery charge), but this is not fully explored in the presented results.
  • Future Work:
    • The authors suggest that this method has great potential for dealing with saturations in various contexts, particularly in power-limited systems where dynamic saturation adjustments could be beneficial.
  • Further exploration of dynamic saturation bounds based on system state is a natural extension.

Summary updated

BibTeX
@article{galloway2015torque,
  author = {Kevin S. Galloway and Koushil Sreenath and Aaron D. Ames and Jessy W. Grizzle},
  title = {{Torque Saturation in Bipedal Robotic Walking Through Control Lyapunov Function-Based Quadratic Programs}},
  journal = {IEEE Access},
  year = {2015},
  doi = {10.1109/access.2015.2419630},
}
Figure