Introduction
In the design of modern autonomous systems and robotics, adaptive sampling has emerged as a critical technique for balancing performance with resource efficiency. By sampling at high frequencies during complex maneuvers and reducing the rate during steady-state operations, systems can optimize for both precision and power consumption. However, this flexibility introduces a significant theoretical challenge: ensuring safety during the transition between different sampling rates.
Traditional safety guarantees, often formulated through control invariant sets, are typically calculated for a fixed sampling interval. This paper addresses the safety gap that occurs when these intervals change, introducing the framework of $M$-step hold control invariance to provide mathematically rigorous safety maps for adaptive-sampling controllers.
The Problem: The Sampling Rate Safety Gap
A control invariant set is a region of the state space where a controller can keep a system indefinitely without violating constraints. A common, yet dangerous, assumption is that a set deemed safe at a slow sampling rate is automatically safe at a faster one.
In reality, the relationship between sampling time and invariance is non-trivial. Switching sampling frequencies mid-operation can inadvertently push a system into a state that is "safe" under one regime but "unsafe" under another. This discrepancy arises because discrete-time models often ignore inter-sample behavior—the state of the system between the "ticks" of the controller's clock. To bridge this gap, a framework is needed that accounts for the continuous-time evolution of the system over varying hold durations.
Methodology: $M$-Step Hold Invariance
The authors propose the $M$-step hold as a solution. In a standard digital control setup, an input is held constant for a single sampling interval ($T_s$). In the $M$-step hold framework, the controller maintains a constant input for $M$ consecutive intervals.
1. $M$-step Precursor Sets
The core of this approach is the redefined "precursor set." Unlike standard definitions that only consider the state at the next sampling instance, the $M$-step precursor set ensures that if a constant input is applied for $M$ steps, every intermediate state remains within the designated safe region. This explicitly accounts for inter-sample constraint satisfaction.
2. Iterative Computation
The paper introduces two primary algorithms to compute the maximal invariant sets:
- Algorithm 1 (Nominal): Designed for systems without external disturbances.
- Algorithm 2 (Robust): Designed for systems facing uncertainty, incorporating a time-varying disturbance bound ($W$). This bound accounts for linearization errors and discretization inaccuracies that accumulate over the $M$ steps.
Key Insights: The "Safety Onion"
The most significant theoretical contribution of this work is the proof of the nesting property: $C^{M+1}_\infty \subseteq C^M_\infty$. This property demonstrates that the maximal invariant set for a slower update rate (larger $M$) is always a subset of the invariant set for a faster update rate (smaller $M$).
This creates a hierarchical structure of safety—a "safety onion." This nesting provides a clear logic for adaptive sampling:
- Increasing Frequency: If a system is safe at a slow sampling rate (inside a smaller, more restrictive set), it is guaranteed to be safe if the controller switches to a faster sampling rate.
- Decreasing Frequency: To safely slow down the sampling rate, the controller must first maneuver the system into the smaller, more restrictive invariant set associated with the target $M$.
Experimental Validation
The framework was validated using two benchmark dynamical systems:
- Constrained Double Integrator: This linear system demonstrated how invariant sets shrink as the hold time $M$ increases, confirming the nesting property visually through polytopic regions.
- Nonlinear Sine-based System: Using the system $\dot{x} = \sin(x) + u$, the authors applied their robust framework. By using Taylor series expansions to bound the error between the actual nonlinear dynamics and a simplified linear model, they successfully calculated robust $M$-step sets that guaranteed safety for the real-world system.
Limitations and Future Work
While providing a rigorous foundation for adaptive sampling, the framework has several limitations:
- Computational Cost: The calculation of these sets involves complex polytope projections. As the dimensionality of the system or the value of $M$ increases, the number of half-planes required to define the sets can grow significantly.
- Conservatism: The robust bounds, particularly for nonlinear systems, can be conservative. Large uncertainty sets may cause the robust invariant set to vanish, even if a safe control strategy exists.
- Policy Design: The paper focuses on defining the safe sets rather than the switching policy. Future research is required to develop the logic that determines exactly when to switch sampling rates based on environmental factors or task demands while remaining within these safety boundaries.
Conclusion
By establishing the $M$-step hold control invariance framework, this research provides a vital link between discrete-time control and continuous-time safety. The nesting property offers a mathematically sound roadmap for designing adaptive controllers that can change their sampling frequency on the fly without compromising the safety of the system.
Summary updated