Fast Trajectory Optimization for Legged Robots Using Vertex-Based ZMP Constraints

Alexander W. Winkler, Farbod Farshidian, Diego Pardo, Michael Neunert, Jonas Buchli

Legged locomotion presents a fundamental control challenge: robots can only exert force by pushing against the environment. Traditional solutions often fall into two categories: computationally heavy full-body trajectory optimization (TO) or fast but rigid hierarchical methods that require pre-planned footholds. This work introduces a middle ground—a fast TO framework that optimizes body motion and footholds simultaneously using a novel vertex-based representation for Zero-Moment-Point (ZMP) constraints.

The Vertex-Based Representation

The core innovation lies in how the Center of Pressure (CoP) is constrained within the support polygon. Instead of traditional linear inequalities, the authors represent the CoP as a weighted sum of the vertices of the feet currently in contact. This barycentric-like approach, utilizing weights ($\lambda$), allows the optimizer to handle point, line, and area contacts uniformly.

The vertex-based representation defines the support area $\mathcal{P}$ and the CoP through weighted vertices ($\lambda$), allowing the optimizer to treat various contact types identically.

By treating footholds as decision variables, the "support area" itself becomes dynamic. This coupling allows the robot to "discover" stabilizing steps, such as stepping outward to recover from a push, rather than following a fixed, pre-calculated gait pattern. To enhance stability, the authors use a robustness cost that penalizes the CoP for moving toward the edges of the feet, encouraging safer motions without the rigidity of hard stability margins.

System Architecture and Optimization

The framework utilizes the Linear Inverted Pendulum (LIP) model, assuming a constant Center of Mass (CoM) height. The continuous problem is converted into a discrete Nonlinear Programming (NLP) problem via direct transcription. To ensure smooth motion, the CoM trajectory is modeled with quartic splines, while foot swing trajectories use cubic polynomials.

The control pipeline integrates the high-level NLP trajectory optimizer with a low-level whole-body controller based on inverse dynamics.

As shown in the system architecture, the high-level NLP solver provides reference trajectories ($q_{ref}$) to a low-level controller. This controller uses inverse dynamics to calculate the necessary joint torques ($\tau$) for the physical robot, bridging the gap between simplified model optimization and complex hardware execution.

Experimental Results

The solver demonstrates exceptional performance, generating a 4-step trot in 0.01 to 0.04 seconds. Even complex 16-step sequences are resolved in under 0.6 seconds, making it suitable for near-real-time applications. The framework was validated on the 80kg HyQ quadruped, successfully executing walking, trotting, and pacing gaits in both simulation and hardware experiments.

Limitations and Future Directions

While highly effective on flat terrain, the current LIP-based model assumes constant foothold heights. Future work aims to extend the formulation to 3D terrain and incorporate the optimizer into a full Model Predictive Control (MPC) loop for real-time disturbance rejection. Additionally, while foothold locations are optimized, the contact sequence (the order of steps) remains predefined in the current implementation.

Summary updated

BibTeX
@article{winkler2017fast,
  author = {Alexander W. Winkler and Farbod Farshidian and Diego Pardo and Michael Neunert and Jonas Buchli},
  title = {{Fast Trajectory Optimization for Legged Robots Using Vertex-Based ZMP Constraints}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2017},
  doi = {10.1109/lra.2017.2723931},
}
Figure