TL;DR
This paper introduces a novel Model Predictive Control (MPC) framework, called Π-MPC, that achieves asymptotically perfect tracking of periodic trajectories, even with significant model mismatches. By augmenting a standard MPC with a periodic disturbance observer, the system can precisely follow repeating paths, a common challenge in robotics.
What Problem Does This Paper Address?
Model Predictive Control (MPC) is a powerful technique for controlling systems, especially those with constraints. However, its performance heavily relies on the accuracy of the predictive model. When the real system's behavior deviates from the model (model mismatch), significant tracking errors can occur, degrading performance and reliability. While complex models can mitigate this, they increase computational burden and design complexity. This paper tackles the problem of achieving perfect tracking for periodic trajectories, a common scenario in robotics (e.g., legged locomotion, industrial manipulation, autonomous racing), despite these model inaccuracies, without resorting to overly complex models.
Methods
The core innovation is the combination of a standard MPC with a periodic disturbance observer. Here's a breakdown:
- Periodic Disturbance Observer: The authors propose a linear observer that estimates "lifted" periodic disturbances. This observer essentially learns the systematic deviations of the real system from the nominal model over a full period. The observer's design is based on an augmented system model that includes both the system state and the estimated periodic disturbance.
- Augmented Model: The nominal linear time-invariant (LTI) model is augmented to include the periodic disturbance. This allows the observer to estimate not just the system state but also the unmodeled dynamics.
- Target Computation: The estimates from the observer are used to compute "targets" for the system's state and input. These targets are designed to achieve perfect tracking if the disturbance estimate were perfect.
- MPC Formulation: A standard MPC formulation is then used, but instead of tracking a reference directly, it tracks these computed targets. The key is that the MPC problem is designed to minimize deviations from these targets, which are themselves designed to compensate for the estimated periodic disturbance.
- Generalization to Nonlinear Systems: The authors also show how this approach can be extended to nonlinear systems, proposing a nonlinear observer and a modified MPC formulation that avoids explicit target computation, making it more practical.
Datasets
Not discussed in the paper. The experiments utilize simulated data from a soft robot model and real-world data from a miniature race car.
Key Insights and Ideas
- Periodic Disturbances are Key: The paper highlights that for periodic tasks, the unmodeled dynamics often exhibit a periodic nature. Exploiting this periodicity is more efficient than trying to model every detail or using general-purpose data-driven methods.
- Observer as a "Periodic Compensator": The periodic disturbance observer acts as a sophisticated compensator, learning and correcting for systematic errors over time. This is more powerful than a simple constant disturbance observer (like in offset-free MPC) for dynamic, repeating trajectories.
- Asymptotic Perfect Tracking: The theoretical guarantee of asymptotically perfect tracking (error going to zero as time goes to infinity) is a significant result, achieved with a relatively simple observer and MPC structure.
- Simplicity and Modularity: The proposed Π-MPC framework is designed to be modular, allowing it to be "plugged in" to existing MPC implementations with minimal changes, which is a major practical advantage.
Key Results
The paper demonstrates the effectiveness of Π-MPC through simulations and hardware experiments:
- Soft Robot Simulation: For a high-dimensional soft robot (nearly 10,000 states) tracking a figure-eight trajectory, Π-MPC achieved asymptotically perfect tracking, reducing the peak error to below 1x10⁻² mm after 50 periods. This is a dramatic improvement over standard MPC (peak error of 46.8 mm) and offset-free MPC (peak error of 28.7 mm).
- Miniature Race Car Experiments: On a real-world race car, Π-MPC reduced peak tracking errors to 2.91 cm after 16 laps, compared to 14.23 cm for a baseline MPC. This shows practical applicability and significant error reduction even with a simple kinematic bicycle model.
Limitations and Future Work
- Non-Deterministic Effects: While the theory guarantees asymptotic perfect tracking, real-world experiments can still exhibit small fluctuations due to non-deterministic effects like noise and delays.
- Non-Periodic Problems: The current framework is specifically designed for periodic trajectories. Extending it to general, non-periodic problems remains an open area for future work.
- Observer Design Complexity: While the observer is simpler than a full system model, designing it for complex nonlinear systems can still be challenging.
- Computational Cost: While the authors claim modularity and efficiency, the computational cost of solving the MPC problem and computing targets might still be a bottleneck for extremely fast or high-dimensional systems.
Summary updated