Safe, Real-Time Active Model Discrimination and Fault Diagnosis for Nonlinear Systems via Differentiable Reachability

arXiv:2606.19590 · cs.RO, cs.SY, eess.SY · Submitted 2026-06-17 · 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: We're looking at the paper: "Safe, Real-Time Active Model Discrimination and Fault Diagnosis for Nonlinear Systems via Differentiable Reachability".

Paper summary: Rosa: So we're looking at this paper, "Safe, Real-Time Active Model Discrimination and Fault Diagnosis for Nonlinear Systems via Differentiable Reachability," and essentially what it claims is a safe way to figure out which model of a system you are actually running in real time, even when things are uncertain.

Dev: Right. The core thesis here is formulating this as an output-feedback, time-varying policy optimization problem that has two main goals: first, robustly enforcing state input safety constraints over a finite time horizon, and second, driving the system to produce sampled measurements that are consistent with only one of the possible models.

Taro: That sounds really ambitious for continuous time nonlinear systems where you have process and measurement disturbances involved. So why is this necessary?

Rosa: Because in these safety-critical settings, if you rely on a controller that trusts a biased measurement or commands a degraded actuator, things can go wrong very quickly. You don't just want to detect that something's broken; you need to isolate which specific fault mode is active fast enough to allow for reconfiguration or fail-safe behavior.

Dev: Exactly. The problem gets tricky because many different candidate models can look very similar when they produce the same measurement signals, like a degraded leg appearing just like a healthy one when it's walking straight. So, you need a way to deliberately excite the system to make those signals distinct and therefore diagnose the fault.

Taro: So the naive question for someone listening is, how do you actually achieve this discrimination in real time without getting bogged down by impossibly complex calculations?

Rosa: That's where they tackle the intractability of exact reachable-set computation for nonlinear systems, which usually makes set-based fault diagnosis methods restricted to linear or affine dynamics, or requiring expensive offline computation. This paper addresses that by using interval over-approximations of reachable state and output sets to make things tractable.

Dev: They use mixed monotonicity in interval analysis to efficiently compute these over-approximations, which is a clever way to get sound bounds on the functions even when you have noise and uncertainty involved. They define an interval approximation of the output reachable set using the state reachable set and some input and measurement noise intervals.

Paper summary: Taro: So if we're talking about what they claim they can compute, it sounds like they are approximating the true reachable sets to make the problem solvable within a real-time framework. What is their main objective for this approximation?

Rosa: The objective is to encode model discrimination as a differentiable objective that penalizes the overlap between the reachable output sets of all possible models. They define this using a quantity called Jq, which involves finding the minimum time interval over which there's an overlap between two models' output reachable sets.

Dev: That overlap penalty is then used in an optimization problem where they seek a policy that minimizes this objective while respecting the safety constraints and the consistency requirement for at most one model. They solve this online using gradient-based methods powered by JAX and differentiable reachability primitives.

Taro: That sounds like a complex, multistep optimization process running on a loop. How do they handle the fact that exact reachable-set computation is NP-hard for general functions?

Rosa: They introduce an output-anticipating refinement scheme to reduce conservatism by conditioning future computations on ambiguity-relevant states. This involves refining the state set used for forward propagation by focusing only on those states that are most relevant to distinguishing between models when computing a control sequence.

Dev: That refinement leads to a multistep optimization problem where they maintain separate intervals for each model pair, which lets them focus their computational effort where it matters most for the diagnosis. It integrates the interval reachability set propagation, the overlap-volume discrimination objective, and those robust safety constraints all into one formulation.

Taro: So if I were driving a robot and needed to diagnose a fault quickly, this approach sounds like it could provide formal guarantees on safety while simultaneously trying to pinpoint exactly which fault is occurring. But what about testing it out in the real world?

Paper summary: Rosa: They've validated this approach on high-dimensional robotic case studies, including a simulated quadrotor, a hardware differential-drive robot, and even quadrupedal navigation systems like the Go2. The results show they achieve reliable model discrimination in under fifty milliseconds for these complex systems.

Dev: And they showed that for the ADMIRE fighter jet model specifically, that output-anticipating refinement discriminates models faster than the uninformed method, and on the Go2 quadrupedal navigation, solving that optimization with a collision avoidance constraint yields a controller with guaranteed obstacle avoidance in forty point two milliseconds plus or minus twenty-five point two microseconds.

Taro: Forty point two milliseconds sounds fast enough for real-time control loops, especially when you're dealing with complex systems. So what's the big picture implication for autonomous systems?

Rosa: It means that we can move toward having controllers that are not just reacting to immediate sensor readings, but actively designing inputs or policies to probe the system in a way that forces it to reveal which model is true, all while keeping safety constraints firmly in place.

Dev: The paper also points out a limitation they acknowledge: for general functions h, computing the inverse of that function's reachable set is NP-hard. So, they suggest that developing fast set inversion algorithms might be an open research topic if we want to push this further.

Taro: So it's a promising framework for handling uncertainty in nonlinear systems where you need both safety and accurate fault isolation at high speed. What are the implications for how we design these real-time diagnostic tools?

Rosa: It suggests that future work could look at assigning a different controller to each model pair, instead of optimizing one shared control law across all pairs, which would be a natural next step to reduce conservatism further.

Dev: I think the main implication is proving that formal safety guarantees can coexist with the need for active, real-time model discrimination in environments where you have process and measurement disturbances.

Taro: So we've covered the summary, why it matters, and what it means for future work in this paper on Safe, Real-Time Active Model Discrimination and Fault Diagnosis for Nonlinear Systems via Differentiable Reachability.

Conclusion: Rosa: So we’re wrapping up our look at this paper, "Safe, Real-Time Active Model Discrimination and Fault Diagnosis for Nonlinear Systems via Differentiable Reachability." Basically, they’ve put together a safe way to figure out which model of a system you're running in real time.

Dev: Yeah, it’s about taking those complex nonlinear systems and process disturbances and giving us a real-time algorithm that handles both safety constraints and fault diagnosis simultaneously.

Taro: What I find interesting is how they tackle the difficulty of exactly calculating reachable sets for nonlinear systems, which usually just takes forever on a computer.

Rosa: They use interval over-approximations, which means they create bounds around where the system could possibly be at any given time and input setting.

Dev: That lets them use JAX and differentiable reachability primitives to solve a policy optimization problem online, which is key for real-time performance.

Taro: And they introduce this output-anticipating refinement strategy, which really helps shrink the uncertainty when they're looking for the right model.

Rosa: So, the authors show that this approach works on high-dimensional systems like simulated quadrotors and even hardware robots, achieving discrimination in under fifty milliseconds.

Dev: The safety case here is strong because it enforces state input constraints robustly while trying to drive the system toward measurements consistent with only one model.

Taro: It’s a big step because it moves beyond just detecting a fault to actively designing control inputs that force the system to reveal which fault mode it’s in.

Rosa: And they also pointed out that for general functions, calculating the inverse of those reachable sets is NP-hard, suggesting that we still have some hard problems left to solve.

Dev: So what this means for us on the ground is a framework where you can actively probe a system’s faults with formal safety guarantees without waiting days for offline analysis.

Taro: It opens up possibilities for autonomous systems that need to react instantly and safely when things go wrong in unpredictable ways.

Institute of Robotics and Intelligent Machines, Georgia Institute of Technology · Department of Aerospace Engineering, University of Illinois Urbana-Champaign

cs.RO, cs.SY, eess.SY

Submitted: 2026-06-17

Updated: 2026-06-17

Journal ref: IEEE Robotics and Automation Letters, vol. 11, no. 12, pp. 14764-14771, Dec. 2026

DOI: 10.1109/LRA.2026.3747586

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 99/100

The gist: Safe, real-time active model discrimination and fault diagnosis for uncertain nonlinear systems with process and measurement disturbances is presented through a safe…

Key concepts

Differentiable Reachability
This technique computes safe, over-approximated sets of all possible states and outputs reachable from a given starting point under system dynamics, input constraints, and noise. By making these computations differentiable using JAX, the method can optimize policies that respect safety limits while simultaneously tracking model consistency.
Output-Anticipating Refinement
This is a strategy used during optimization where the algorithm focuses on 'ambiguity-relevant states' when computing future control steps. By refining the state set based on which models are most likely to be correct, it reduces conservatism and allows for faster discrimination between different fault scenarios.
Model Discrimination Objective
This is a differentiable objective function designed to measure how much two possible system models overlap in their reachable output sets. Minimizing this overlap forces the optimized control policy to steer the system's actual outputs away from the regions predicted by incorrect fault models, thereby isolating the true model.
Interval Over-approximations
Since exact reachability computation for nonlinear systems is too complex, this method uses interval analysis to create conservative bounds (over-approximations) for state and output sets. These bounds guarantee that the true reachable set is entirely contained within the computed interval, ensuring safety guarantees even with uncertainty.

Terminology

Summary

Safe, real-time active model discrimination and fault diagnosis for uncertain nonlinear systems with process and measurement disturbances is presented through a safe, real-time algorithm that uses differentiable reachability to enforce safety constraints while driving the system toward sampled measurements consistent with at most one model The resulting optimization is solved efficiently online with gradient-based methods using JAX and differentiable reachability primitives

Problem Formulation

The paper formulates active fault diagnosis as an output-feedback, time-varying policy optimization problem that couples robust safety with a diagnosability requirement over sampled outputs The feasibility problem (3) requires finding a policy π that ensures state–input safety constraints are robustly satisfied under every candidate fault model and uncertainty realization This problem is difficult because exact reachable-set computation for nonlinear systems is intractable, necessitating the use of interval over-approximations of reachable state and output sets

Tractable Reachability Computation

To make the problem tractable, the authors compute interval over-approximations of reachable state and output sets They leverage mixed monotonicity to efficiently compute these over-approximations using interval analysis and immrax Definition 3 defines the Interval Approximation of Output Reachable Set as Ry(T, Rx, u, [v]):= [Hm(Rx(T), u, [v]), Hm(Rx(T), u, [v])] when it is subject to measurement noise v ∈ [v] and input u over all t ∈ [0, T]

Discrimination Objective

The model-discrimination constraint (3f) is encoded as a differentiable objective J(π) that penalizes overlap between the reachable output sets of possible models This objective is defined as J(π):= X q∈Q Jq(π) where Jq(π):= min t∈Tˆ Vol m∈q Ym(t, π(t, y))! The volumetric overlap Vol used in (8) can be substituted by many other metrics that can measure separation between two compact sets, such as Hausdorff distance [28]

Output-Anticipating Refinement

The paper introduces an output-anticipating refinement scheme (9) to reduce conservatism by conditioning future computation on ambiguity-relevant states This involves refining the state set used for forward propagation by focusing on ambiguity-relevant states, which is achieved when computing a control sequence The resulting multistep optimization problem integrates interval reachableset propagation, the overlap-volume discrimination objective, and robust safety constraints within a single formulation

Experimental Validation

Experiments on high-dimensional robotic case studies demonstrate that the approach achieves reliable model discrimination in under 50 ms The method was evaluated on a simulated quadrotor, a hardware differential-drive robot, and quadrupedal navigation systems For the ADMIRE fighter jet model, the output-anticipating refinement discriminates models faster than the uninformed method The Go2 quadrupedal navigation showed that solving (8) with a collision avoidance constraint yields a controller with guaranteed obstacle avoidance in 40.2 ms ± 25.2 µs

Conclusion and Future Work

The paper concludes by presenting a safe, real-time framework for active model discrimination and fault diagnosis in uncertain nonlinear systems A key contribution is the output-anticipating refinement strategy (9), which reduces conservatism by propagating ambiguity-relevant state intervals A natural next step is to explore this idea by allowing a different controller to be assigned to each model pair, rather than optimizing a shared control law across all pairs The paper highlights that the method can handle input and state constraints, as seen in the Go2 case study The runtime comparison shows that the output-anticipating refinement (9) discriminates models faster than uninformed method (8) The authors note that for general functions h, computing H−1(y) is NP-hard, suggesting fast setinversion algorithms are an open research topic The paper's results show that the approach demonstrates fast runtimes and reliable discrimination performance for both sensor and actuator fault scenarios The final section suggests exploring a different controller to be assigned to each model pair

REFERENCES

arXiv:2606.19590v1 [cs. p. 0

Where the paper says it

arXiv:2606. 19590v1 [cs.

RO] 17 Jun 2026

--- Page 1 ---

Safe, Real-Time Active Model Discrimination and Fault Diagnosis for Nonlinear Systems via Differentiable Reachability Xinpei Ni Melkior Ornik Glen Chou Samuel Coogan Abstract— We present a safe, real-time algorithm for active fault diagnosis and model discrimination for uncertain continuous-time nonlinear systems with process and measurement disturbances. Given a finite set of candidate models representing nominal and faulty modes, including actuator and sensor faults, we formulate an output-feedback, time-varying policy optimization problem that (i) robustly enforces stateinput safety constraints over a finite horizon and (ii) drives the system to produce sampled measurements consistent with at most one model, enabling deterministic diagnosis. p. 0

Where the paper says it

Safe, Real-Time Active Model Discrimination and Fault Diagnosis for Nonlinear Systems via Differentiable Reachability Xinpei Ni1 Melkior Ornik2 Glen Chou1 Samuel Coogan1 Abstract— We present a safe, real-time algorithm for active fault diagnosis and model discrimination for uncertain continuous-time nonlinear systems with process and measurement disturbances. Given a finite set of candidate models representing nominal and faulty modes, including actuator and sensor faults, we formulate an output-feedback, time-varying policy optimization problem that (i) robustly enforces stateinput safety constraints over a finite horizon and (ii) drives the system to produce sampled measurements consistent with at most one model, enabling deterministic diagnosis.

To solve this problem in real time, we develop a tractable approximation using interval over-approximations of reachable state and output sets, and encode diagnosability via a differentiable objective that penalizes overlap between the reachable output sets of possible models. p. 0
Where the paper says it

Given a finite set of candidate models representing nominal and faulty modes, including actuator and sensor faults, we formulate an output-feedback, time-varying policy optimization problem that (i) robustly enforces stateinput safety constraints over a finite horizon and (ii) drives the system to produce sampled measurements consistent with at most one model, enabling deterministic diagnosis. To solve this problem in real time, we develop a tractable approximation using interval over-approximations of reachable state and output sets, and encode diagnosability via a differentiable objective that penalizes overlap between the reachable output sets of possible models.

The resulting optimization is solved efficiently online with gradient-based methods using JAX and differentiable reachability primitives. p. 0
Where the paper says it

To solve this problem in real time, we develop a tractable approximation using interval over-approximations of reachable state and output sets, and encode diagnosability via a differentiable objective that penalizes overlap between the reachable output sets of possible models. The resulting optimization is solved efficiently online with gradient-based methods using JAX and differentiable reachability primitives.

We evaluate our method on sensor and actuator fault diagnosis (up to 11 fault modes) in several highdimensional nonlinear robotic systems, including a simulated quadrotor and fighter-jet model, a hardware differential-drive robot, and quadrupedal navigation. p. 0
Where the paper says it

The resulting optimization is solved efficiently online with gradient-based methods using JAX and differentiable reachability primitives. We evaluate our method on sensor and actuator fault diagnosis (up to 11 fault modes) in several highdimensional nonlinear robotic systems, including a simulated quadrotor and fighter-jet model, a hardware differential-drive robot, and quadrupedal navigation.

Across these case studies, our approach achieves reliable model discrimination in under 50 ms, outperforming baselines in discrimination success rate and speed while providing formal safety guarantees

--- Page 2 ---

I. INTRODUCTION Faults are inevitable in long-duration robotic deployments. p. 0

Where the paper says it

edu 2Melkior Ornik is with the Department of Aerospace Engineering, University of Illinois Urbana-Champaign, Urbana, IL 61801, USA mornik@illinois. INTRODUCTION Faults are inevitable in long-duration robotic deployments.

For instance, sensors can drift or saturate, actuators gradually lose effectiveness, and unmodeled dynamics can emerge as hardware wears. p. 0
Where the paper says it

INTRODUCTION Faults are inevitable in long-duration robotic deployments. For instance, sensors can drift or saturate, actuators gradually lose effectiveness, and unmodeled dynamics can emerge as hardware wears.

In safety-critical settings, delayed or incorrect diagnosis can be catastrophic: controllers that trust biased measurements or command degraded actuators may quickly violate safety constraints or destabilize the system. p. 0
Where the paper says it

For instance, sensors can drift or saturate, actuators gradually lose effectiveness, and unmodeled dynamics can emerge as hardware wears. In safety-critical settings, delayed or incorrect diagnosis can be catastrophic: controllers that trust biased measurements or command degraded actuators may quickly violate safety constraints or destabilize the system.

Thus, robots must not only detect that a fault has occurred, but also isolate which fault mode is valid quickly enough to enable reconfiguration or fail-safe behavior. p. 0
Where the paper says it

In safety-critical settings, delayed or incorrect diagnosis can be catastrophic: controllers that trust biased measurements or command degraded actuators may quickly violate safety constraints or destabilize the system. Thus, robots must not only detect that a fault has occurred, but also isolate which fault mode is valid quickly enough to enable reconfiguration or fail-safe behavior.

A standard formalization of fault isolation is model discrimination [1]: given a finite set of candidate dynamical models (e. p. 0
Where the paper says it

Thus, robots must not only detect that a fault has occurred, but also isolate which fault mode is valid quickly enough to enable reconfiguration or fail-safe behavior. A standard formalization of fault isolation is model discrimination [1]: given a finite set of candidate dynamical models (e.

g., a nominal model and multiple fault models), determine which model is consistent with observed behavior. p. 0
Where the paper says it

A standard formalization of fault isolation is model discrimination [1]: given a finite set of candidate dynamical models (e. , a nominal model and multiple fault models), determine which model is consistent with observed behavior.

This task is difficult because diagnosis must be performed from noisy, partial measurements under disturbances and uncertainty. p. 0
Where the paper says it

, a nominal model and multiple fault models), determine which model is consistent with observed behavior. This task is difficult because diagnosis must be performed from noisy, partial measurements under disturbances and uncertainty.

Moreover, many distinct candidate models can produce similar measurement signals. p. 0
Where the paper says it

This task is difficult because diagnosis must be performed from noisy, partial measurements under disturbances and uncertainty. Moreover, many distinct candidate models can produce similar measurement signals.

For instance, a quadruped with a degraded leg may appear similar to a healthy quadruped with a biased IMU when commanded to walk straight. p. 0
Where the paper says it

Moreover, many distinct candidate models can produce similar measurement signals. For instance, a quadruped with a degraded leg may appear similar to a healthy quadruped with a biased IMU when commanded to walk straight.

Control synthesis, however, can simplify diagnosis: for the quadruped, a carefully chosen turning maneuver may generate sensor measurements that disambiguate the failure mode. p. 0
Where the paper says it

For instance, a quadruped with a degraded leg may appear similar to a healthy quadruped with a biased IMU when commanded to walk straight. Control synthesis, however, can simplify diagnosis: for the quadruped, a carefully chosen turning maneuver may generate sensor measurements that disambiguate the failure mode.

This motivates active fault diagnosis (AFD): rather than passively monitoring outputs, we design control inputs or policies that deliberately excite the system to make px(m) (m) 1 2 3 4 Tim 5 e (s) py (m) m mm m Original Output-reachable Intervals Nominal vs Actuator Fault After Refinement Nomina l Actuator Fault Sensor Fault Intersection Trajectory (Actuator Fault,α=0. p. 0
Where the paper says it

This motivates active fault diagnosis (AFD): rather than passively monitoring outputs, we design control inputs or policies that deliberately excite the system to make 0. 5 px (m) 1 2 3 4 Tim 5 e (s) 0.

25)(A) Illustration of output reachable intervals with and without sensor anticipating refinement. p. 0
Where the paper says it

25) py (m) (A) (B) (C) Fig. 1: (A) Illustration of output reachable intervals with and without sensor anticipating refinement.

Using output-anticipating refinement allows the intersections of output reachable intervals to shrink faster and enables faster diagnosis. p. 0
Where the paper says it

1: (A) Illustration of output reachable intervals with and without sensor anticipating refinement. Using output-anticipating refinement allows the intersections of output reachable intervals to shrink faster and enables faster diagnosis.

(B) Top view of a hardware experiment in which a separating output-feedback controller is computed by (9) and applied on a unicycle system. p. 0
Where the paper says it

Using output-anticipating refinement allows the intersections of output reachable intervals to shrink faster and enables faster diagnosis. (B) Top view of a hardware experiment in which a separating output-feedback controller is computed by (9) and applied on a unicycle system.

p. 0
Where the paper says it

Using output-anticipating refinement allows the intersections of output reachable intervals to shrink faster and enables faster diagnosis. (B) Top view of a hardware experiment in which a separating output-feedback controller is computed by (9) and applied on a unicycle system.

The controller successfully discriminates a sensor fault from an actuator fault in 5 seconds while ensuring that the system’s true position avoids an obstacle shaded in gray. p. 0
Where the paper says it

(B) Top view of a hardware experiment in which a separating output-feedback controller is computed by (9) and applied on a unicycle system. The controller successfully discriminates a sensor fault from an actuator fault in 5 seconds while ensuring that the system’s true position avoids an obstacle shaded in gray.

p. 0
Where the paper says it

(B) Top view of a hardware experiment in which a separating output-feedback controller is computed by (9) and applied on a unicycle system. The controller successfully discriminates a sensor fault from an actuator fault in 5 seconds while ensuring that the system’s true position avoids an obstacle shaded in gray.

While the output reachable set for sensor fault (in green) overlaps with the obstacle, this is caused by the sensor fault’s output corruption and the robot’s true state i. p. 0
Where the paper says it

The controller successfully discriminates a sensor fault from an actuator fault in 5 seconds while ensuring that the system’s true position avoids an obstacle shaded in gray. While the output reachable set for sensor fault (in green) overlaps with the obstacle, this is caused by the sensor fault’s output corruption and the robot’s true state i.

p. 0
Where the paper says it

The controller successfully discriminates a sensor fault from an actuator fault in 5 seconds while ensuring that the system’s true position avoids an obstacle shaded in gray. While the output reachable set for sensor fault (in green) overlaps with the obstacle, this is caused by the sensor fault’s output corruption and the robot’s true state i.

e. position remains clear of the obstacle at all times. p. 0
Where the paper says it

While the output reachable set for sensor fault (in green) overlaps with the obstacle, this is caused by the sensor fault’s output corruption and the robot’s true state i. position remains clear of the obstacle at all times.

p. 0
Where the paper says it

While the output reachable set for sensor fault (in green) overlaps with the obstacle, this is caused by the sensor fault’s output corruption and the robot’s true state i. position remains clear of the obstacle at all times.

(C) Angled view of a hardware experiment where a separating controller is computed by (8) and applied on a quadruped to diagnose actuator and sensor faults in two separate runs. p. 0
Where the paper says it

position remains clear of the obstacle at all times. (C) Angled view of a hardware experiment where a separating controller is computed by (8) and applied on a quadruped to diagnose actuator and sensor faults in two separate runs.

p. 0
Where the paper says it

position remains clear of the obstacle at all times. (C) Angled view of a hardware experiment where a separating controller is computed by (8) and applied on a quadruped to diagnose actuator and sensor faults in two separate runs.

Frames corresponding to the actuator fault are tinted in red, and frames corresponding to the sensor fault are tinted green. p. 0
Where the paper says it

(C) Angled view of a hardware experiment where a separating controller is computed by (8) and applied on a quadruped to diagnose actuator and sensor faults in two separate runs. Frames corresponding to the actuator fault are tinted in red, and frames corresponding to the sensor fault are tinted green.

p. 0
Where the paper says it

(C) Angled view of a hardware experiment where a separating controller is computed by (8) and applied on a quadruped to diagnose actuator and sensor faults in two separate runs. Frames corresponding to the actuator fault are tinted in red, and frames corresponding to the sensor fault are tinted green.

The controller successfully diagnoses the fault in 2 seconds

--- Page 3 ---

III. PROBLEM FORMULATION We consider a set of M uncertain nonlinear continuous-time dynamical systems fm and measurement output maps hm x˙ m(t) = fm(xm(t), u(t), w(t), θa), (1a) ym(t) = hm(xm(t), u, v, θo), (1b) indexed by m ∈ M:= 0,. p. 2

Where the paper says it

PROBLEM FORMULATION We consider a set of M uncertain nonlinear continuous-time dynamical systems fm and measurement output maps hm x˙ m(t) = fm(xm(t), u(t), w(t), θa), (1a) ym(t) = hm(xm(t), u(t), v(t), θo), (1b) indexed by m ∈ {0, . , M −1}, with state xm ∈ X , control u ∈ U, process noise w ∈ W, measurement output y ∈ Y, and measurement noise v ∈ V.

.., M −1, with state xm ∈ X, control u ∈ U, process noise w ∈ W, measurement output y ∈ Y, and measurement noise v ∈ V. p. 2
Where the paper says it

PROBLEM FORMULATION We consider a set of M uncertain nonlinear continuous-time dynamical systems fm and measurement output maps hm x˙ m(t) = fm(xm(t), u(t), w(t), θa), (1a) ym(t) = hm(xm(t), u(t), v(t), θo), (1b) indexed by m ∈ {0, . , M −1}, with state xm ∈ X , control u ∈ U, process noise w ∈ W, measurement output y ∈ Y, and measurement noise v ∈ V.

Here (f0, h0), i.e., m = 0 denotes the nominal model which defines the state evolution and measurements of the healthy system The remaining models (fm, hm), for m ∈ M−1 denote various fault modes such as actuator fault or sensor fault Definition 1. p. 2
Where the paper says it

, m = 0, denote the nominal model which defines the state evolution and measurements of the healthy system, with parameters θ ∗ a ∈ Θa and θ ∗ o ∈ Θo. The remaining models (fm, hm), for m ∈ {1, .

Sampled measurement set. p. 0
Where the paper says it

Safe, Real-Time Active Model Discrimination and Fault Diagnosis for Nonlinear Systems via Differentiable Reachability Xinpei Ni1 Melkior Ornik2 Glen Chou1 Samuel Coogan1 Abstract— We present a safe, real-time algorithm for active fault diagnosis and model discrimination for uncertain continuous-time nonlinear systems with process and measurement disturbances. Given a finite set of candidate models representing nominal and faulty modes, including actuator and sensor faults, we formulate an output-feedback, time-varying policy optimization problem that (i) robustly enforces stateinput safety constraints over a finite horizon and (ii) drives the system to produce sampled measurements consistent with at most one model, enabling deterministic diagnosis.

We assume that over a time interval [0, T], measurements are collected from the system at a discrete set of Nt pre-determined time instants Tˆ:= tˆ1,. p. 2
Where the paper says it

Sampled measurement set. We assume that over a time interval [0, T], measurements are collected from the system at a discrete set of Nt pre-determined time instants Tˆ := {tˆ1, .

..,tˆNt subset of [0, T]. p. 2
Where the paper says it

We assume that over a time interval [0, T], measurements are collected from the system at a discrete set of Nt pre-determined time instants Tˆ := {tˆ1, . ,tˆNt } ⊂ [0, T].

Assuming the system is operating under the true model m∗ ∈ M, this yields the sampled measurement set Yˆ:= ym∗ (t̂i) i=1. p. 2
Where the paper says it

Assuming the system is operating under the true model m∗ ∈ {0, . , M − 1}, this yields the sampled measurement set Yˆ := {yˆi := ym∗ (tˆi)} Nt i=1.

Let Yˆ denote the set of all possible Yˆ that can be realized under different disturbances w(·) and v(·). p. 2
Where the paper says it

, M − 1}, this yields the sampled measurement set Yˆ := {yˆi := ym∗ (tˆi)} Nt i=1. Let Yˆ denote the set of all possible Yˆ that can be realized under different disturbances w(·) and v(·).

Definition 2. p. 2
Where the paper says it

, fm ̸= f0 and hm ̸= h0. Definition 1.

Consistency set. p. 2
Where the paper says it

Definition 2. Consistency set.

We define I(Yˆ) as the set of models consistent with the sampled measurement set Yˆ as I(Yˆ):= m ∈ M ∃x(0) ∈ X0, ∀w(·), v(·), s. p. 2
Where the paper says it

Consistency set. We define I(Yˆ ) as the set of models consistent with the sampled measurement set Yˆ as I(Yˆ ) := {m ∈ {0, .

t. ym(t̂i) = ˆyi, ∀i We consider the following fault modes: Actuator fault: where there is a deviation in the dynamics function fm (x, u, w, θa), i. p. 2
Where the paper says it

, M − 1}, denote various fault modes. We denote the set of systems as M := {(fm, hm)} M−1 m=0 and consider the following fault modes: • Actuator fault: where there is a deviation in the dynamics function fm(x, u, w, θa), i.

e., fm!= f0 and hm = h0 Sensor fault: where there is a deviation in the output function hm (x, u, v, θo), i. p. 2
Where the paper says it

We denote the set of systems as M := {(fm, hm)} M−1 m=0 and consider the following fault modes: • Actuator fault: where there is a deviation in the dynamics function fm(x, u, w, θa), i. , fm ̸= f0 and hm = h0, • Sensor fault: where there is a deviation in the output function hm(x, u, v, θo), i.

e., fm = f0 and hm!= h0 Simultaneous fault: where there are deviations in both the dynamics and output, i. p. 2
Where the paper says it

, fm ̸= f0 and hm = h0, • Sensor fault: where there is a deviation in the output function hm(x, u, v, θo), i. , fm = f0 and hm ̸= h0, • Simultaneous fault: where there are deviations in both the dynamics and output, i.

e., fm!= f0 and hm!= h0

--- Page 4 ---

IV. PRELIMINARIES A. p. 2

Where the paper says it

IV and then describe our approximate solution method for (3) in Sec. PRELIMINARIES A.

State and Output Reachable Sets Given (fm, hm) for some model m ∈ M as defined in (1) and an initial state set X0, we define its state reachable set at time T as: Xm T:= x(T) ∈ X x(0) ∈ X0, x˙ m = fm(xm, u, w), u(τ) ∈ U, w(τ) ∈ W, ∀τ ∈ [0, T) Similarly, the output reachable set is defined as the image of the state reachable set under the observation map hm: Ym T:= hm(x, u, v) x ∈ Xm T, u ∈ π(t, Xt), v ∈ V Because computing (4) and (5) exactly is difficult, we exploit mixed monotonicity to efficiently compute their overapproximations For the above system, we leverage interval analysis to obtain inclusion functions Fm = [Fm, Fm], Hm = [Hm, Hm] over a given interval [x] = [x, x], [v] = [v, v], [w] = [w, w] such that Fm([x], u, v) ≤ fm(x) ≤ Fm([x], u, v]) and Hm([x], u, w]) ≤ hm(x) ≤ Hm([x], u, w]) for all x ∈ [x], v ∈ [v], w ∈ [w] This set of inclusion functions induces an embedding system (Fm, Hm) for the system (fm, hm)

--- Page 5 ---

B. Approximating Reachable Sets with immrax To efficiently compute such embedding systems, we turn to immrax [26] a parallelizable and differentiable JAXbased Python toolbox that facilitates interval analysis and mixed monotone reachability through composable function transforms immrax leverages core features of JAX, such as just-in-time (JIT) compilation, GPU acceleration, and automatic differentiation to efficiently compute sound reachable set over-approximations Definition 3. p. 3

Where the paper says it

Approximating Reachable Sets with immrax To efficiently compute such embedding systems, we turn to immrax [26], a parallelizable and differentiable JAXbased Python toolbox that facilitates interval analysis and mixed monotone reachability through composable function transforms. immrax leverages core features of JAX, a Python package for high-performance computation, such as just-in-time (JIT) compilation, GPU acceleration, and automatic differentiation, to efficiently compute sound reachable set over-approximations.

Interval Approximation of Output Reachable Set Given a system model like (1), a time horizon T, inclusion functions Fm for fm and Hm for hm, and an interval overapproximation of state reachable set at time T, Rx(T), a corresponding interval approximation of its output reachable set at time T is defined as Ry(T, Rx, u, [v]):= [Hm(Rx(T), u, [v]), Hm(Rx(T), u, [v])] when it is subject to measurement noise v ∈ [v] and input u over all t ∈ T. p. 3
Where the paper says it

Interval Approximation of Output Reachable Set. Given a system model like (1), a time horizon T, inclusion functions Fm for fm and Hm for hm, and an interval overapproximation of state reachable set at time T, Rx(T), a corresponding interval approximation of its output reachable set at time T is defined as Ry(T, Rx, u, [v]) := [Hm(Rx(T), u, [v]), Hm(Rx(T), u, [v])] when it is subject to measurement noise v ∈ [v] and input u over all t ∈ [0, T].

Proposition 1. p. 3
Where the paper says it

Given a system model like (1), a time horizon T, inclusion functions Fm for fm and Hm for hm, and an interval overapproximation of state reachable set at time T, Rx(T), a corresponding interval approximation of its output reachable set at time T is defined as Ry(T, Rx, u, [v]) := [Hm(Rx(T), u, [v]), Hm(Rx(T), u, [v])] when it is subject to measurement noise v ∈ [v] and input u over all t ∈ [0, T]. Proposition 1.

Ry as given in Definition 3 overapproximates the true output-reachable set Y; i. p. 3
Where the paper says it

Proposition 1. Ry as given in Definition 3 overapproximates the true output-reachable set Y; i.

e., Y ⊆ Ry

--- Page 6 ---

V. METHOD We now describe our method, which conservatively solves (3). p. 3

Where the paper says it

□ Relative to (3), robust safety constraint satisfaction is enforced via the reachable set over-approximations Rx, and the consistency constraint (3f) is relaxed to a cost penalty. METHOD We now describe our method, which conservatively solves (3).

First, we show how to encode the model-discrimination constraint (3f) as a differentiable objective (Sec. p. 3
Where the paper says it

METHOD We now describe our method, which conservatively solves (3). First, we show how to encode the model-discrimination constraint (3f) as a differentiable objective (Sec.

V-A) Using a tractable over-approximation of the state and output reachable sets to obtain an optimization problem that can be implemented efficiently in JAX (Sec. p. 3
Where the paper says it

Using a tractable over-approximation of the state and output reachable sets (Sec. IV-B) to obtain an optimization problem that can be implemented efficiently in JAX (Sec.

V-B) We now present an optimization problem for active fault diagnosis min u J(u):= X q∈Q Jq(u) (8a) s. p. 4
Where the paper says it

In particular, it integrates interval reachableset propagation, the overlap-volume discrimination objective, and robust safety constraints within a single formulation. We now present an optimization problem for active fault diagnosis.

t. Jq(uk) = min tk∈Tˆ Vol m∈q [y]m q (tk) (8b), [x]m(tk+1) = Rm x (tk+1, [x] m(tk), uk, [w]) (8c), [y]m(tk) = Hm([x] m(tk), uk, [v]) = Hm([x] m(tk), uk, [v]) (8d) Rm x (t) ⊆ S

--- Page 7 ---

D. Output-Anticipating Refinement for Tighter Model Discrimination Starting from (8) we refine the state set used for forward propagation by focusing on ambiguity-relevant states To further reduce conservatism, we maintain a separate interval for each model pair q ∈ Q The resulting multistep optimization problem is: for any time step k ∈ 0, 1, …, K, where K denotes the horizon or the number of control steps we wish to solve for, the optimal control uk is found as: min uk∈U J(uk) = X q∈Q(tk) Jq(uk) (9a) s. p. 4

Where the paper says it

Output-Anticipating Refinement for Tighter Model Discrimination Starting from (8), we refine the state set used for forward propagation by focusing on ambiguity-relevant states. Noticing that finding a controller tsatisfying (6) is equivalent to separating each pair at any time when an output is measured, we argue that when computing a control sequence, only states that will lead to unseparated outputs require further separation.

t. Jq(uk) = min tk∈Tˆ Vol m∈q [y]m q (tk) (9b), [x]m(tk+1) = (8c), [y]m(tk) = Hm([x] m(tk), uk, [v]) (9c), ˜[y] m q(tk) = min m∈q [y]m q (tk) (9d), ˜[x] m q(tk) = H−1m (˜[y] qm(tk)) ∩ [x]q m(tk) (9e), [x]m(tk+1) = Rm x (t˜x] m q(tk), uk, [w]) (9f), [x]m(τ) ⊆ S

--- Page 8 ---

VI. RESULTS We present hardware and simulation experiments of our proposed active model discrimination scheme on three different systems Unicycle shows (9) can run in milliseconds on a physical nonlinear system with a small set of models ADMIRE demonstrates (9) can be applied on a highdimensional nonlinear system with many candidate models and solves at real-time rates Go2 shows (8) is a valid choice for sensor fault models involving unobserved states while enforcing safety The solution to (8) results in full separation between output intervals at T = 2s

--- Page 4 ---

Fig. 1: (A) Illustration of output reachable intervals with and without sensor anticipating refinement Using output-anticipating refinement allows the intersections of output reachable intervals to shrink faster and enables faster diagnosis. p. 0

Where the paper says it

1: (A) Illustration of output reachable intervals with and without sensor anticipating refinement. Using output-anticipating refinement allows the intersections of output reachable intervals to shrink faster and enables faster diagnosis.

(B) Top view of a hardware experiment in which a separating output-feedback controller is computed by (9) and applied on a unicycle system. The controller successfully discriminates a sensor fault from an actuator fault in 5 seconds while ensuring that the system’s true position avoids an obstacle shaded in gray. While the output reachable set for sensor fault (in green) overlaps with the obstacle, this is caused by the sensor fault’s output corruption and the robot’s true state i.e. position remains clear of the obstacle at all times. (C) Angled view of a hardware experiment where a separating controller is computed by (8) and applied on a quadruped to diagnose actuator and sensor faults in two separate runs. Frames corresponding to the actuator fault are tinted in red, and frames corresponding to the sensor fault are tinted green. The controller successfully diagnoses the fault in 2 seconds

--- Page 5 ---

Fig. 4: Hardware validation of active fault diagnosis on Go2 Table II: Runtime comparison between our method and baseline Ours Baseline Unicycle Open Loop 3. p. 7

Where the paper says it

TABLE II: Runtime comparison between our method and baseline. Task/Run Time Ours Baseline Unicycle Open Loop 3.

84ms solver fails Unicycle Output Feedback with Collision Avoidance 47. p. 7
Where the paper says it

Task/Run Time Ours Baseline Unicycle Open Loop 3. 84ms solver fails Unicycle Output Feedback with Collision Avoidance 47.

2ms - ADMIRE Open Loop 45. p. 7
Where the paper says it

84ms solver fails Unicycle Output Feedback with Collision Avoidance 47. 2ms - ADMIRE Open Loop 45.

3ms 301s Go2 Open Loop 40. p. 7
Where the paper says it

2ms - ADMIRE Open Loop 45. 3ms 301s Go2 Open Loop 40.

2ms 22.6s

--- Page 8 ---

Fig.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI can perform safe, real-time active fault diagnosis for uncertain continuous-time nonlinear systems by formulating fault isolation as an output-feedback, time-varying policy optimization problem that enforces stateinput safety constraints over a finite horizon.

  2. Improved AI can achieve deterministic diagnosis by driving the system to produce sampled measurements consistent with at most one model through a differentiable objective that penalizes overlap between the reachable output sets of possible models.

  3. Improved AI can solve this optimization problem efficiently online using gradient-based methods with JAX and differentiable reachability primitives, enabling fast replanning as candidate models are pruned online in under 50 ms.

  4. Improved AI can utilize interval over-approximations of reachable state and output sets to compute tractable approximations, which allows for the enforcement of robust safety constraints even for high-dimensional nonlinear systems where exact set computation is intractable.

  5. Improved AI can employ an output-anticipating refinement scheme that reduces conservatism by conditioning future computation on ambiguity-relevant states, leading to tighter model discrimination in hardware experiments like the ADMIRE fighter jet model.

Abstract

We present a safe, real-time algorithm for active fault diagnosis and model discrimination for uncertain continuous-time nonlinear systems with process and measurement disturbances. Given a finite set of candidate models representing nominal and faulty modes, including actuator and sensor faults, we formulate an output-feedback, time-varying policy optimization problem that (i) robustly enforces state-input safety constraints over a finite horizon and (ii) drives the system to produce sampled measurements consistent with at most one model, enabling deterministic diagnosis. To solve this problem in real time, we develop a tractable approximation using interval over-approximations of reachable state and output sets, and encode diagnosability via a differentiable objective that penalizes overlap between the reachable output sets of possible models. The resulting optimization is solved efficiently online with gradient-based methods using JAX and differentiable reachability primitives. We evaluate our method on sensor and actuator fault diagnosis (up to 11 fault modes) in several high-dimensional nonlinear robotic systems, including a simulated quadrotor and fighter-jet model, a hardware differential-drive robot, and quadrupedal navigation. Across these case studies, our approach achieves reliable model discrimination in under 50 ms, outperforming baselines in discrimination success rate and speed while providing formal safety guarantees.

Sources

Related papers