⚡ electrical engineering

Staggered Integral Online Conformal Prediction for Safe Dynamics Adaptation with Multi-Step Coverage Guarantees

This paper proposes Staggered Integral Online Conformal Prediction (SI-OCP), a novel algorithm that utilizes an integral score function to provide long-run coverage guarantees for quantifying model uncertainty in adaptive systems without state derivative measurements, thereby enabling safe dynamics adaptation when synthesized with robust safety-critical controllers like tube MPC.

Original authors: Daniel M. Cherenson, Dimitra Panagou

Published 2026-04-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Daniel M. Cherenson, Dimitra Panagou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are driving a self-driving car through a city where the weather changes every second, and the road surface is slippery and unpredictable. Your car has a "brain" (a computer model) that tries to guess what will happen next.

The Problem: The "Blind Spot" of Safety
Usually, to drive safely, the car needs to know exactly how much it might slide or drift.

The Old Way: Engineers would say, "Let's assume the worst possible slide ever!" This makes the car drive very slowly and cautiously, like a turtle, because it's terrified of crashing.
The New Way (Adaptive Learning): The car tries to learn as it drives. If it starts to slide, it updates its brain to say, "Oh, it's slippery here, I'll adjust."
The Catch: To know exactly how much it slid, the car needs to measure its speed changing instantly (like a derivative). But in the real world, sensors are often too slow or noisy to give that instant measurement. Without that instant data, the car can't be sure if its new "brain" is actually working or if it's just guessing wrong. If it guesses wrong, it might crash.

The Solution: SI-OCP (The "Delayed Report Card" System)
The authors of this paper, Daniel and Dimitra, created a clever system called Staggered Integral Online Conformal Prediction (SI-OCP).

Here is how it works, using a simple analogy:

1. The "Integral Score" (The Receipt, Not the Item)

Since the car can't see the instant slip, it looks at the receipt of the trip instead.

Instead of asking, "How much did I slip right now?" (which is impossible to measure), the system asks, "How far off was my total path from where I thought I would go over the last 5 seconds?"
This is like checking your bank account at the end of the day. You might not know exactly what you spent on a specific coffee at 2:03 PM, but you know your total balance is off by $5. That total difference tells you something important about your spending habits.
In the paper, they call this an Integral Score. It sums up all the little errors over a short window of time to give a clear picture of the total "messiness" of the environment.

2. The "Staggered" Approach (The Relay Race)

Here is the tricky part: You can only calculate that "total messiness" after the 5 seconds are over. But the car needs to know the safety margin before it drives the next 5 seconds.

The Analogy: Imagine a relay race where the runners are passing a baton.
- Runner A runs the first 5 seconds.
- Runner B starts running the next 5 seconds immediately.
- By the time Runner B finishes their leg, Runner A is done and can finally say, "Hey, I was off by 2 meters!"
- Runner B takes that information and says, "Okay, for the next runner, I'll assume a 2-meter error."
The system does this with multiple threads (like multiple runners) running in a staggered pattern. While one thread is driving the car, another thread is analyzing the data from the previous drive to update the safety rules for the future.

3. The "Conformal Prediction" (The Safety Bubble)

Once the system calculates how "messy" the last few seconds were, it uses a statistical trick called Conformal Prediction.

Think of this as drawing a safety bubble (or a tube) around the car's path.
If the last trip was very bumpy, the bubble gets big. The car drives slowly and stays far away from obstacles to be safe.
If the last trip was smooth, the bubble gets small. The car can drive faster and squeeze through narrow gaps between obstacles because it knows it won't slide much.
The magic is that this bubble is mathematically guaranteed to be big enough 90% (or 99%) of the time, even if the weather changes unpredictably. It doesn't need to know why the wind is blowing; it just knows how much the car actually moved.

The Real-World Test: The Drone in the Wind

The authors tested this on a quadcopter drone flying through a forest with trees (obstacles) and strong, changing winds.

The Drone's Brain: It used a complex AI (a Deep Neural Network) to learn the wind patterns on the fly.
The Result:
- Without this system: The drone was too scared to fly through a narrow gap between two trees because it assumed the wind would be terrible.
- With SI-OCP: The drone realized, "Hey, the wind was actually pretty calm in the last few seconds." It shrank its safety bubble just enough to squeeze through the narrow gap safely, while still having a mathematical guarantee that it wouldn't crash.
- Even if the AI was bad: They tested a scenario where the AI was frozen (not learning). The system noticed the AI was failing, made the safety bubble huge, and the drone simply refused to fly through the gap, staying safe instead of crashing.

Summary

This paper gives robots a way to be brave but smart.

It allows them to learn and adapt to new, scary environments (like a drone in a storm).
It doesn't need perfect sensors to measure every tiny movement.
It uses a "delayed report card" system to constantly update a dynamic safety bubble.
This ensures the robot stays safe 99% of the time, even when the world is chaotic and the robot's own brain is still learning.

In short: It's a safety net that gets tighter or looser based on how well the robot is actually doing, ensuring it never falls through.

1. Problem Statement

The paper addresses the challenge of safe control for adaptive systems operating in uncertain, dynamic environments. Specifically, it targets continuous-time systems where:

Uncertainty is Unknown: The system dynamics include unmodeled disturbances ( $\Delta$ ) and modeling errors.
Adaptation is Online: The controller uses an adaptive model (e.g., a Deep Neural Network) to learn these dynamics in real-time.
Measurement Limitations: While the system state $x(t)$ is measurable, the state derivative $\dot{x}(t)$ is not. This prevents the direct calculation of the instantaneous residual error (the difference between the true dynamics and the learned model) required by standard uncertainty quantification methods.
Safety Requirements: Safety-critical controllers (like Model Predictive Control) require multi-step coverage guarantees. They need to bound the accumulated error over a future prediction horizon to ensure recursive feasibility and safety, rather than just providing instantaneous (one-step) bounds.

The Core Gap: Existing Online Conformal Prediction (OCP) methods require ground-truth labels for prediction errors immediately after a prediction is made. In adaptive dynamics, the "label" (the true residual) is unavailable online because it requires $\dot{x}$ . Furthermore, standard OCP provides only one-step guarantees, which are insufficient for controllers that plan over a finite horizon.

2. Methodology: Staggered Integral Online Conformal Prediction (SI-OCP)

The authors propose SI-OCP, a distribution-free framework that quantifies uncertainty without needing state derivatives or assuming specific disturbance distributions.

A. Integral Non-Conformity Score

Instead of measuring instantaneous error, SI-OCP measures the integrated effect of the disturbance over a rolling time window $[t_k - T_p, t_k]$ .

The Score ( $S_k$ ): Defined as the supremum of the integral difference between the actual state trajectory and the trajectory predicted by the nominal model (including the adaptive parameter estimate $\hat{\theta}$ ):
$S_k = \sup_{\tau_1, \tau_2 \in [t_k-T_p, t_k]} \left\| x(\tau_2) - x(\tau_1) - \int_{\tau_1}^{\tau_2} f_{nom}(s, u(s), \hat{\theta}(s)) ds \right\|$
Significance: This score effectively captures the lumped effect of approximation errors, adaptation errors, and exogenous disturbances using only measurable states and inputs.

B. Staggered Multi-Step Updates

To provide multi-step coverage (guarantees over a horizon $T_p$ ), the algorithm employs a "staggered" approach:

Thread Partitioning: The prediction horizon is divided into $P$ discrete intervals. The algorithm maintains $P$ independent "threads" (conformal thresholds).
Delayed Labeling: At time $t_k$ , the algorithm computes the score $S_k$ based on the past horizon. This score serves as the "ground truth" label for the prediction made $T_p$ seconds ago (by thread $j = k \mod P$ ).
Staggered Update: The conformal threshold $q^{(j)}_k$ for the active thread is updated using sub-gradient descent on the pinball loss. Because the threads are decoupled, the system can update thresholds continuously while maintaining valid coverage guarantees for the future horizon.

C. From Integral to Point-wise Bounds

The integral score $S_k$ is converted into a point-wise disturbance bound $d_k$ for the upcoming horizon. This conversion relies on the Lipschitz constant of the disturbance ( $L_d$ ) and the prediction horizon length ( $T_p$ ).

If the integral bound is small, the point-wise bound is derived from the square root of the integral (Case 1).
If the integral bound is large, a linear relationship is used (Case 2).
This dynamic bound $d_k$ is fed into a Robust Tube Model Predictive Control (RT-MPC) or similar safety-critical controller to ensure the system remains within a safe set.

3. Key Contributions

Derivative-Free Uncertainty Quantification: The introduction of an integral non-conformity score allows for uncertainty quantification in adaptive systems without requiring unmeasurable state derivatives ( $\dot{x}$ ).
Multi-Step Coverage Guarantees: The staggered update mechanism extends OCP from instantaneous guarantees to asymptotic long-run multi-step coverage. This ensures that the accumulated error over a prediction horizon is bounded with a probability of at least $1-\alpha$ .
General Applicability: The framework is agnostic to the underlying adaptation law. It works with linear-in-parameter models and complex all-layer Deep Neural Network (DNN) adaptations.
Safety Integration: The method provides a rigorous theoretical proof that integrating SI-OCP with a robust predictive controller yields asymptotic long-run safety (the system stays in the safe set with probability $1-\alpha$ ).

4. Simulation Results

The authors validated SI-OCP using a 3D quadcopter navigating through obstacles under time-varying wind and unmodeled aerodynamic drag.

Setup: The adaptive model was an all-layer DNN updated via Self-Supervised Meta Learning (SSML). The controller was a Dynamic Tube MPC (DTMPC).
Scenario 1 (Adaptation Enabled): The DNN learned the dynamics, reducing the residual error. SI-OCP produced a tight uncertainty bound, allowing the MPC to generate a narrow "tube" that successfully navigated through a tight gap between obstacles.
- Coverage: The true disturbance was bounded 98.77% of the time (target was 90%).
Scenario 2 (Adaptation Disabled): The DNN was frozen with poor initial weights. The residual error was high. SI-OCP correctly detected this, expanding the uncertainty bound.
- Result: The MPC generated a wider tube, preventing the quadcopter from attempting to fly through the narrow gap, thereby avoiding a collision.
- Coverage: The bound held 94.0% of the time.

5. Significance

This work bridges a critical gap between data-driven adaptive control and formal safety guarantees.

Overcoming Conservatism: Traditional safe adaptive control often relies on worst-case bounds that are overly conservative, limiting performance. SI-OCP provides data-driven, tight bounds that allow for higher performance while maintaining safety.
Enabling Complex Learning: By removing the need for state derivatives and handling non-i.i.d. (non-exchangeable) data, SI-OCP enables the safe deployment of complex, non-linear learning models (like deep neural networks) in safety-critical robotics.
Robustness to Distribution Shift: The method is distribution-free, making it robust to changing environments where the nature of the disturbance is unknown or evolves over time.

In summary, SI-OCP provides a mathematically rigorous, practical method for ensuring that adaptive robots can learn from their environment in real-time without sacrificing safety guarantees.