Perfecting Periodic Trajectory Tracking: Model Predictive Control with a Periodic Observer (Π-MPC)

Luis A. Pabon, Johannes Köhler, John Irvin Alora, Patrick Benito Eberhard, Andrea Carron, Melanie N. Zeilinger, Marco Pavone

Introduction

Model Predictive Control (MPC) has become a cornerstone of modern robotics, valued for its ability to handle complex nonlinearities and state constraints by predicting future system behavior. However, the performance of MPC is fundamentally limited by the accuracy of its underlying mathematical model. When a "model mismatch" occurs—where the simplified physics used by the controller fails to capture the true dynamics of the system—persistent tracking errors inevitably emerge.

While contemporary research often attempts to bridge this gap using high-capacity, data-driven models like Deep Neural Networks or Gaussian Processes, these approaches introduce significant computational overhead and implementation complexity. Researchers from Stanford and ETH Zurich have proposed a more elegant alternative: Π-MPC. By leveraging the inherent periodicity of many robotic tasks—such as walking, industrial manufacturing cycles, or lap-based racing—Π-MPC achieves near-perfect tracking using remarkably simple models.

The Problem: The Cost of Model Mismatch

In standard MPC, a controller solves an optimization problem at every time step to find the best sequence of inputs. If the model predicts a certain movement but the physical robot encounters unmodeled friction, soft-body deformation, or aerodynamic drag, the robot will drift from its intended path. Standard "offset-free" MPC techniques attempt to correct this by estimating a constant disturbance, but these are often insufficient for dynamic, time-varying errors that occur throughout a periodic trajectory.

Methodology: The Periodic Disturbance Observer

The core innovation of Π-MPC is the integration of a Periodic Disturbance Observer within the MPC framework. This approach combines the strengths of Repetitive Control (RC) with the constraint-handling capabilities of MPC. The system operates through four primary stages:

  1. The "Lifted" Model: Rather than estimating a single static offset, the observer utilizes a "lifted" disturbance vector. This vector represents the specific model mismatch at every discrete time step within a single period $N$.
  2. Luenberger Observer: A linear estimator updates these disturbance estimates in real-time. It compares the predicted output to the actual measurement, refining the disturbance map for each point in the cycle with every passing period.
  3. Target Computation: The system calculates the specific state and input "targets" required to cancel out the estimated periodic disturbances. This ensures that the reference the MPC tracks is adjusted to account for known errors.
  4. MPC Integration: The MPC optimizes control inputs to track these adjusted targets. Because the observer learns the pattern of the error, the MPC can preemptively compensate for disturbances before they occur.

Crucially, the researchers provided a theoretical proof demonstrating that this architecture achieves asymptotic zero tracking error. This means that regardless of the complexity of the unmodeled dynamics, the error will eventually vanish as the system learns the periodic pattern.

Experimental Validation

The researchers evaluated Π-MPC across two distinct and challenging platforms to test its versatility:

Soft Robot Simulation

Soft robots are notoriously difficult to model due to their infinite-dimensional state space. Using the SOFA framework, the team simulated a "Diamond" soft robot with a high-fidelity Finite Element Method (FEM) model consisting of 9,768 dimensions. Despite this complexity, the Π-MPC was configured with a drastically simplified linear model of only 6 states.

In a figure-8 tracking task, the baseline MPC suffered an average tracking error of 30.5 mm. After 50 periods of learning, Π-MPC reduced this error to just 0.004 mm—a 5000x reduction. In contrast, standard offset-free MPC only reached 18.6 mm, as it could not account for the time-varying nature of the errors within the figure-8.

Miniature Race Car Hardware

To test the system in the real world, the team used a 1:28 scale RC car navigating a complex track. The controller utilized a simple kinematic bicycle model, which ignores complex tire dynamics and surface interactions. Π-MPC reduced the peak tracking error from 14.2 cm (baseline) to 2.9 cm, representing a 5x improvement in precision on physical hardware.

Key Insights and Future Directions

The success of Π-MPC highlights a powerful principle in robotics: Simplicity is sufficient if the observer is smart. You do not need a 10,000-state model to control a complex soft robot if your observer can learn the repeating pattern of the model's failures. Furthermore, the modular design of Π-MPC allows it to be implemented as an "add-on" to existing MPC frameworks, making it highly practical for industrial applications.

Limitations

Currently, Π-MPC assumes a strict periodicity with a known, constant period $N$. It is not yet optimized for stochastic or non-repeating disturbances. Additionally, the memory requirements for the "lifted" disturbance vector scale linearly with the period length, which could pose challenges for high-frequency sampling over very long cycles.

Future Work

Future research will explore extending this framework to non-periodic problems and developing methods to handle varying period lengths—such as a race car that changes its lap time as it gains speed. By bridging the gap between simple models and complex reality, Π-MPC provides a robust path toward high-precision autonomous systems.

Summary updated

BibTeX
@inproceedings{pabon2024perfecting,
  author = {Luis A. Pabon and Johannes Köhler and John Irvin Alora and Patrick Benito Eberhard and Andrea Carron and Melanie N. Zeilinger and Marco Pavone},
  title = {{Perfecting Periodic Trajectory Tracking: Model Predictive Control with a Periodic Observer (Π-MPC)}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2024},
  doi = {10.1109/iros58592.2024.10801564},
}
Figure