Probabilistic Reachable Set Estimation for Saturated Systems with Unbounded Additive Disturbances

arXiv:2504.03443 · math.OC, cs.SY, eess.SY · Submitted 2025-04-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Probabilistic Reachable Set Estimation for Saturated Systems with Unbounded Additive Disturbances".

Dev: In this paper, an analytical approach is presented for synthesizing ellipsoidal probabilistic reachable sets (PRS) of saturated systems subject to unbounded additive noise,

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: So, we’re looking at the paper titled "Probabilistic Reachable Set Estimation for Saturated Systems with Unbounded Additive Disturbances," which sounds like it's about using convex optimization to build these ellipsoidal probabilistic reachable sets for systems dealing with unbounded noise and input saturation. Dev I see. The authors are tackling a problem that’s harder than the standard bounded noise cases because they're dealing with inputs that can hit hard limits, modeled as saturations.

Taro: From an autonomy research standpoint, I wonder how this handles scenarios where the system faces unexpected disturbances or sudden changes in environment dynamics that aren't captured by a fixed covariance model. Rosa That’s a valid point, Taro; if we want true robustness for autonomous systems, we need to know if these probabilistic bounds can cope with situations outside the strict assumptions they laid out.

Dev: The core idea they present is computing a contraction factor for the saturating error dynamics to get tight bounds on the evolution of the system's state uncertainty. Taro And that contraction factor calculation seems central, but I need to know how that relates to actual control loop performance and latency, because those are huge factors in real-time systems.

Rosa: The implication here is that for systems where you have hard input constraints modeled as direct saturation on the control input, this analytical approach provides a way to construct reachable sets without relying solely on the worst-case scenarios. Dev So it’s about getting a better probabilistic guarantee by using optimization methods to find a better contraction rate than just looking at the open-loop or closed-loop rates separately.

Taro: If they can effectively compute that effective contraction rate, does it mean we can design systems that are inherently more resilient to those unbounded disturbances, even if the exact nature of the noise is unknown? Rosa That’s a big question for deployment; if this method works robustly, it could be key for safety-critical applications where probabilistic guarantees matter.

The paper's summary: Dev: To summarize what we just touched on about "Probabilistic Reachable Set Estimation for Saturated Systems with Unbounded Additive Disturbances," the paper outlines an analytical method using convex optimization to compute a contraction factor for the saturating error dynamics. Rosa It’s essentially a way to tightly bound how the error evolves in these saturated systems, which then allows them to construct accurate ellipsoidal probabilistic reachable sets.

Taro: So, they are defining these probabilistic reachable sets based on a sequence of sets R k where the probability of being inside that set stays above some violation level epsilon as time progresses. Dev That definition seems standard for stochastic processes, but the paper’s innovation is how they build those specific sets when the system dynamics involve those saturations.

Rosa: They introduce a framework where they consider linear systems affected by independent, zero-mean noise and hard input constraints modeled as direct saturation on the control input. Taro And they leverage established results from saturated systems theory to bound the saturated error dynamics within a convex set whose vertices encode all possible saturation scenarios—fully saturated, unsaturated, and partially saturated configurations.

Dev: That construction of a set whose vertices cover all possible saturation states is what lets them compute quadratic Lyapunov functions that are valid for the error dynamics. Rosa And then they figure out an effective contraction rate that sits between the worst-case open-loop rate and the best-case unsaturated closed-loop rate.

Taro: That intermediate rate sounds promising because it acknowledges both the potential for instability in open-loop operation and the performance gains from closing the loop, which is exactly what we need when dealing with uncertain environments.

The paper's improvements: Rosa: The main improvement they present is deriving a tight bound on the expectation of a quadratic transformation on the error, which they use to compute accurate ellipsoidal probabilistic reachable sets with a user-defined violation probability. Dev That means instead of just having some loose worst-case bound, they can generate sets that match the desired level of risk tolerance we set for our control task.

Taro: If you can define the set based on a specific violation probability epsilon, does that mean we move away from relying on overly conservative bounds and toward something more tailored to the specific mission requirements? Rosa Exactly; it allows for a design where you explicitly specify how often you expect the system to violate a constraint, which is much more useful than just knowing it *might* violate it under the absolute worst conditions.

Dev: They also suggest an improvement in system design by enforcing Assumption three which is a compatibility condition, to ensure that closed-loop dynamics are faster than open-loop dynamics in the region of linearity. Taro So they’re not just analyzing the noise; they’re guiding the system design itself to operate optimally within that region of linearity where things behave predictably.

Rosa: And this allows them to optimize control gains specifically for performance while still maintaining stability guarantees, which is a nice balance for any real-world control architecture. Dev It seems like they’re moving from just proving feasibility under constraints to actively shaping the reachable set to meet mission objectives probabilistically.

Conclusion: Dev: So, wrapping up on "Probabilistic Reachable Set Estimation for Saturated Systems with Unbounded Additive Disturbances," the paper successfully synthesizes ellipsoidal probabilistic reachable sets by computing a contraction factor via convex optimization. Rosa The implication is that we can now construct much more accurate and user-defined probabilistic bounds for linear systems under unbounded noise and saturation, which is a step up from what was previously possible.

Taro: I think the real impact here is in how it informs proactive system design; if we can map out these high-risk areas using their resulting maps, we can build smarter autonomous agents that avoid dangerous states before they happen. Dev And from an engineering standpoint, getting that tight bound based on the effective contraction rate gives us a much better handle on the latency and failure modes of the control loop itself.

Rosa: Indeed, it shows how analytical methods can give us concrete tools to quantify uncertainty in complex control scenarios involving saturation and heavy noise. Dev It’s a solid piece of work that moves beyond just checking if a system *can* do something under ideal conditions to actually characterizing its probabilistic performance when things get messy.

Taro: I'm excited to see how this framework scales to nonlinear systems or even more complex disturbance models in future work, though the current focus on linear systems is a necessary starting point. Rosa Absolutely, and I’m eager to see how we can adapt these principles for those more challenging environments we discussed earlier. Dev We'll be keeping an eye on this paper as we look at applying these ideas to our next set of control problems.

Univ. Grenoble Alpes · CNRS

math.OC, cs.SY, eess.SY

Submitted: 2025-04-04

Updated: 2025-08-31

Comments: 11 pages, LaTeX; section III.C framing rephrased, Remark 3 added, numerical example presentation updated, typos corrected, references updated

DOI: 10.1109/CDC57313.2025.11312343

Code: https://github.com/CarloKaram/PRS-Sat-Sy

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: In this paper, an analytical approach is presented for synthesizing ellipsoidal probabilistic reachable sets (PRS) of saturated systems subject to unbounded additive noise, utilizing convex

Key concepts

Probabilistic Reachable Sets (PRS)
These are ellipsoidal sets used to represent the possible states of a system over time while accounting for uncertainty. The paper focuses on constructing these sets accurately when the system has input saturation and unbounded noise.
Contraction Factor
This is a factor computed using convex optimization that bounds how quickly the error in a saturated system evolves. It sits between worst-case and best-case rates, providing an effective contraction rate for analysis.
Input Saturation
This refers to situations where the control input hits hard limits, modeled as direct saturation. The paper uses established theory to bound the error dynamics across all possible saturation configurations.
Violation Probability (epsilon)
This is a user-defined level of risk tolerance used to construct the probabilistic reachable sets. It allows designers to specify how often they expect the system to violate a constraint, moving beyond overly conservative bounds.

Terminology

Summary

In this paper, an analytical approach is presented for synthesizing ellipsoidal probabilistic reachable sets (PRS) of saturated systems subject to unbounded additive noise, utilizing convex optimization methods to compute a contraction factor for the saturating error dynamics, which allows for tight bounding and accurate construction of reachable sets. The proposed method is applicable to independent, zero-mean disturbances with a known covariance.

The problem addresses the synthesis of probabilistic reachable sets (PRS) under unbounded probabilistic noise, which remains an open problem in contrast to well-established methods for bounded noise. The authors present an approach for synthesizing PRS in the context of an indirect feedback formulation where nominal and error dynamics are decoupled. They consider linear systems affected by unbounded, independent, zero-mean noise and subject to hard input constraints modeled as a direct saturation on the control input:

We consider linear systems affected by unbounded, independent, zero-mean noise, and subject to hard input constraints modeled as a direct saturation on the control input.

Leveraging established results from saturated systems theory [18], [19], the authors bound the saturated error dynamics within a well-defined convex set constructed such that its vertices encode all possible saturation scenarios: This set is constructed such that its vertices encode all possible saturation scenarios: fully saturated (open-loop), unsaturated (closed-loop), and partially saturated configurations. This bounding enables the computation of quadratic Lyapunov functions defined on that set, which remain valid for the error dynamics. Motivated by the realization that the system probabilistically switches between saturation modes, they compute its effective contraction rate which lies in between the open-loop (worst-case) and unsaturated closed-loop (best-case) rates. They subsequently derive a tight bound on the expectation of a quadratic transformation on the error, enabling the computation of accurate ellipsoidal probabilistic reachable sets with user-defined violation probability. The authors state: To the best of our knowledge, no constructive PRS design method incorporating the feedback on the error term exists in the nonlinear (saturated) setting.

The theoretical framework involves defining key concepts related to probabilistic reachability:

(Definition 1)

"A sequence of sets Rk are said to be probabilistic reachable sets (PRS) of violation level ε ∈ [0, 1] for a stochastic process involving values in R n if e0 = 0 ⇒ Pr ek ∈ Rk ≥ 1 − ε, ∀k > 0."

(Definition 2)

A set R is said to be a probabilistic ultimate bound (PUB) of violation level ε ∈ [0, 1] for a stochastic process involving values in R n, if ∀e0 ∈ R n, ∃t s.t. Pr ek ∈ R ≥ 1 − ε ∀k ≥ t.

(Definition 3)

"A set R∞ is said to be a probabilistic invariant set (PIS) of violation level ε ∈ [0, 1] for a stochastic process involving values in R n, if e0 ∈ R∞ =⇒ Pr ek ∈ R∞ ≥ 1 − ε ∀k > 0."

The system dynamics are formulated as:

xk+1 = Axk + Buk + wk, (1)

Pr xk ∈ X ≥ 1 − εx ∀k ∈ N0, (2)

uk ∈ U ∀k ∈ N0, (3)

where the input constraint is modeled as a saturation function: xk+1 = Axk + Bφ (uk) + wk, (4) with φi(u) = sign (ui) min (ui, 1) ∀i ∈ N+m. (5)

The system state is split into nominal and error parts:

xk = zk + ek, (6)

uk = vk + Kek, (7)

leading to the error dynamics: ek+1 = f(ek, vk) + wk, (9b)

The core of the synthesis involves bounding the image of the error term through the saturated function. The authors utilize support functions and set inclusion theorems to show that for a given input norm constraint, the image of the nonlinear system is contained within a polytope whose vertices encode all possible saturation scenarios: Theorem 2 states that for all e ∈ R n and any v∞ ≤ 1, the image f(e, v) is contained in a known polytope F(e) whose vertices are given by the linear combination A + P i∈J B(i)Ki, where J ⊆ N+m.

Using this result, they construct a base PRS for the error term. Under specific conditions related to a positive definite matrix P satisfying certain contraction rate properties:

"Proposition 1.

Improvements for AI systems

As a fastidious researcher, I have analyzed the provided paper, Probabilistic Reachable Set Estimation for Saturated Systems with Unbounded Additive Disturbances. This work provides a novel analytical framework for synthesizing ellipsoidal probabilistic reachable sets (PRS) and ultimate bounds (PUB) for linear systems subject to hard input constraints modeled as saturations and unbounded additive noise.

The core contribution is the derivation of an effective, tighter contraction rate—the average of the open-loop rate and the closed-loop rate—by exploiting the system's region of linearity (RL). This leads to a more conservative but ultimately more accurate bound than existing methods that rely solely on worst-case scenarios.

Here are specific improvements for AI systems based on this research, categorized by application:

)

Improved AI System Capabilities:

The paper's methodology is directly applicable to developing robust and safety-critical control architectures for dynamic, real-world AI agents. The improved systems will possess the following capabilities:

  1. textbf Robust Stochastic Model Predictive Control (SMPC) for Autonomous Robotics/Vehicles (Safety & Reliability):

  2. textbf Adaptive Safety Constraint Tightening (Constraint Management):

  3. textbf Enhanced Uncertainty Quantification in Decision Making (Risk Assessment):

)

Improved AI System Capabilities:

  1. Robust Stochastic Model Predictive Control (SMPC) for Autonomous Robotics/Vehicles: The system can compute control inputs that not only satisfy hard physical actuator limits but also guarantee a specified probability of satisfying state constraints (e.g., staying within safe zones, maintaining collision avoidance corridors) despite unbounded, unpredictable sensor or environmental noise.

  2. Adaptive Safety Constraint Tightening: The system will dynamically adjust its safety margins based on the estimated region of linearity (RL). When the agent operates in a region where the control input is not saturated (i.e., near the optimal operating point), it uses a tighter, closed-loop contraction rate to compute more aggressive and accurate reachable set bounds, allowing for faster, safer maneuvers than systems relying on conservative open-loop bounds.

  3. Enhanced Uncertainty Quantification in Decision Making: The system will generate probabilistic confidence intervals (PRS/PUB) around its predicted future states. Instead of a single deterministic trajectory, it provides a quantifiable measure of the likelihood that the system will violate a constraint (e.g., There is a 95% probability that the robot's position will remain within this volume). This allows higher-level AI planners to make risk-aware decisions, trading off performance for guaranteed probabilistic safety levels.

)

Detailed Implementation Specifics:

The improved system can perform the following specific actions:

  1. Compute a Tight, Time-Varying Safety Bound: It will continuously calculate the expected error evolution bound, which is explicitly defined by Proposition 3 as a function of the effective contraction rate (which optimally balances open-loop and closed-loop dynamics). This allows for an online, adaptive safety guarantee that is significantly tighter than bounds derived from purely worst-case analysis.

  2. Distinguish Operational Modes: The system will internally monitor whether it is operating in the region of linearity (RL) versus a saturated regime. It will switch between two distinct bounding strategies (Proposition 3 vs. the conservative bound in (15)) based on this internal state, maximizing performance where possible while maintaining rigorous probabilistic guarantees where necessary.

  3. Optimize Control Gains for Performance: The system can be designed to explicitly enforce Assumption 3 (compatibility condition) to ensure the closed-loop dynamics are faster than the open-loop dynamics in the RL region, enabling a design that maximizes control authority without sacrificing stability guarantees.

  4. Generate Probabilistic Risk Maps: By using the resulting set definitions (Lemma 3 and Lemma 4), it can generate spatial or state-space maps indicating high-risk areas where the probability of constraint violation exceeds a threshold, informing path planning algorithms to avoid those regions proactively.

Sources

Related papers