Abstract
Designing a model predictive control (MPC) scheme that enables a mobile robot to safely navigate through an obstacle-filled environment is a complicated yet essential task in robotics. In this technical report, safety refers to ensuring that the robot respects state and input constraints while avoiding collisions with obstacles despite the presence of disturbances and measurement noise. This report offers a step-by-step approach to implementing nonlinear model predictive control (NMPC) schemes addressing these safety requirements. It focuses specifically on NMPC as a foundation for safe mobile robot navigation, providing a practical path from theoretical concepts to mathematical proofs and implementation.
Introduction
While Nonlinear Model Predictive Control (NMPC) is a powerful tool for robot navigation, a significant gap often exists between theoretical formulations and real-world implementation. In practice, robots encounter disturbances (e.g., wind, uneven terrain) and measurement noise from imperfect sensors. Without accounting for these uncertainties, an NMPC scheme may lead to "infeasible" states—mathematical dead-ends where no safe path forward exists, potentially resulting in collisions. This report serves as a pedagogical bridge, providing a framework to guarantee recursive feasibility (ensuring a safe path always exists) and stability (ensuring the target is reached) under realistic conditions.
Methodology: A Three-Stage Design
The NMPC design is structured into three progressive levels of complexity:
- Tracking MPC (Nominal): This baseline assumes a perfect model. It introduces "terminal ingredients"—a terminal cost, a terminal control law, and a terminal set. These components act as a "safe harbor" at the end of the finite prediction horizon, ensuring the optimization problem remains solvable indefinitely.
- Robust MPC (Disturbances): This stage introduces unknown but bounded external disturbances. The authors employ Constraint Tightening, creating a "buffer zone" or tube around the planned trajectory to account for potential deviations.
- Robust Output-Feedback MPC (Noise + Disturbances): The most advanced case addresses scenarios where the robot's state must be estimated via an observer. A second layer of constraint tightening is added to compensate for estimation errors.
A critical component of this methodology is the use of Control Contraction Metrics (CCMs). By solving Linear Matrix Inequalities (LMIs), a metric is derived that guarantees the robot's actual trajectory will "contract" toward the desired reference, allowing for formal quantification of safety buffer sizes.
Theoretical Insights
- The "Safe Harbor" Concept: NMPC safety depends on the state at the end of the prediction horizon. The "terminal set" defines a region where a simple auxiliary controller can maintain safety indefinitely.
- Dynamic Tightening: Unlike rigid tube MPC, this approach allows safety buffers to grow over the horizon, reflecting increasing uncertainty over time.
- Geometric-Control Integration: The framework converts non-convex collision-free spaces into sequences of convex polytopic constraints, enabling efficient real-time optimization.
- Observer-Aware Safety: Safety margins explicitly incorporate observer error; higher sensor noise necessitates larger safety buffers.
Results and Guarantees
The report provides rigorous mathematical proofs for Recursive Feasibility and Asymptotic Convergence across all three control levels. It offers concrete "design recipes" using Semi-Definite Programs (SDPs) to pre-compute optimal control gains ($K$) and safety tube shapes ($P$). While noise may prevent perfect target tracking, the error is guaranteed to remain within a calculable bound proportional to noise and disturbance levels.
Limitations and Future Work
Despite its strengths, the approach faces challenges:
- Computational Complexity: Solving LMIs and SDPs for high-degree-of-freedom robots can be expensive.
- Conservatism: Polytopic approximations and fixed tubes may lead to overly cautious behavior in narrow passages.
- Dynamic Environments: Current proofs assume moving obstacles can be over-approximated by known sets. Handling highly unpredictable, reactive agents like humans remains a key area for future research.
Summary updated