TL;DR
This paper introduces a hybrid control strategy that combines model-based feedback linearization with robust, data-driven predictive control (DDPC). By using a mechanical model to simplify robot dynamics and data to "mop up" residual errors from model inaccuracies and noise, the authors achieve superior trajectory tracking for robotic manipulators compared to standard model-based or purely data-driven methods.
What Problem Does This Paper Address?
Predictive control is a cornerstone of modern robotics, but it faces a fundamental dilemma. Model Predictive Control (MPC) relies on accurate mathematical models, which are notoriously difficult to obtain for complex systems with uncertain parameters (like friction or exact mass distributions). On the other hand, Data-Enabled Predictive Control (DeePC) skips the model and predicts behavior directly from recorded data. However, purely data-driven methods can be data-hungry and struggle with the inherent nonlinearities of robotic systems.
The authors address this gap by proposing a "model-informed" approach. They use a "good enough" mechanical model to handle the heavy lifting of nonlinear dynamics through feedback linearization, then employ a robust data-driven controller to learn and compensate for the systematic errors (residuals) that the model misses.
Methods
The proposed framework consists of two main stages:
- Feedback Linearization (FL): A mechanical model maps virtual acceleration commands to joint torques. If the model were perfect, each joint would behave like a simple "double integrator" (a linear system where input is acceleration and output is position).
- Geometric Robust DDPC: Since models are rarely perfect, the system exhibits "residual dynamics." The authors use Geometric Robust DDPC to learn these residuals from data. Instead of fitting a single linear model, this method fits a low-dimensional trajectory subspace on a Grassmannian manifold. To handle noise and model mismatch, the controller considers a "ball" of possible subspaces around the fitted one and optimizes for the worst-case scenario within that ball.
This robust approach is illustrated in Figure 1, where the controller seeks a solution that performs well even if the "true" system dynamics lie on an adverse subspace ($S^a$) slightly different from the one fitted to the data ($\hat{S}$).
Datasets
The method was evaluated using simulations of a planar 2R manipulator (a two-joint robotic arm) moving under the influence of gravity. The authors tested several scenarios to challenge the controller:
- Exact vs. Inaccurate Parameters: Testing how the system handles errors in mass and damping.
- Transfer Tests: Evaluating if a controller tuned for one arm ("Arm A") remains effective when the arm's physical properties change ("Arm B") or when sensor noise increases.
- Varying Noise Levels: Testing robustness against measurement noise ($\sigma_y$) ranging from 0.001 to 0.010 radians.
Key Insights and Ideas
- Don't Throw Away the Physics: Purely data-driven control is hard. By using a mechanical model for feedback linearization, the authors simplify the learning task from "learn the whole robot" to "learn the small errors in my model."
- Learning the Residuals: The data-driven part of the controller effectively captures systematic departures from ideal behavior, such as those caused by incorrect mass estimates or unmodeled damping.
- Robustness via Geometry: Treating trajectory subspaces as points on a Grassmannian manifold provides a principled way to define "uncertainty" in a data-driven model. Optimizing for the worst-case subspace within a neighborhood makes the controller significantly more resilient to noise than standard DeePC.
Key Results
- Superior Prediction: When model parameters were inaccurate, the "Learned behavior" (DDPC) was an order of magnitude more accurate than a simple double-integrator model. For example, at low noise, prediction RMSE dropped from 0.2096 to 0.0153 rad.
- Improved Tracking: In trajectory tracking tasks with parameter errors, Geometric DDPC consistently outperformed double-integrator MPC, achieving lower cyclic errors ($E_{cyc}$).
- Resilience to Noise: The geometric approach maintained stable performance and lower error spreads compared to direct DeePC as measurement noise increased, as shown in the transfer tests.
Limitations and Future Work
- Computational Cost: The robust optimization involves a "scenario-exchange" algorithm that is more computationally intensive than the standard quadratic programming used in MPC or DeePC.
- Simulation-Based: While the results are promising, the study is currently limited to high-fidelity simulations. Testing on real robotic hardware is a necessary next step.
- Constraint Handling: The current predictive formulation does not explicitly incorporate joint or torque constraints, which are critical for safe operation in many real-world robotic applications. The authors note that these were checked post-hoc but not enforced during optimization.
Summary updated