💻 computer science

Latent Dynamics-Aware OOD Monitoring for Trajectory Prediction with Provable Guarantees

This paper proposes a latent dynamics-aware OOD monitoring framework for trajectory prediction in safety-critical CPS that models in-distribution error evolution via Hidden Markov Models and extends the cumulative Maximum Mean Discrepancy approach within a quickest changepoint detection formulation to achieve provable guarantees on detection delay and false-alarm rates without requiring explicit knowledge of post-change distributions.

Original authors: Tongfei Guo, Lili Su

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tongfei Guo, Lili Su

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 teaching a self-driving car how to navigate a busy city. You train it on thousands of hours of video from sunny days, clear roads, and polite drivers. The car learns perfectly. But then, you take it out on a rainy Tuesday night in a chaotic construction zone. The car's "brain" (its prediction model) starts making mistakes. It might think a pedestrian is standing still when they are actually running, or it might fail to predict that a car ahead will slam on its brakes.

The problem is: How does the car know right now that it is confused?

Most current systems are like a student taking a test who only realizes they got a question wrong after the teacher grades the whole exam. By then, it's too late to stop the car. This paper proposes a new "safety monitor" that acts like a vigilant co-pilot, watching the car's predictions in real-time to scream, "Hey, something is wrong!" before an accident happens.

Here is a breakdown of their solution, DC-MMD, using simple analogies.

1. The Problem: The "Silent" Failure

In the real world, traffic isn't static. Sometimes the road is smooth (low risk), and sometimes it's a chaotic mess of merging cars and jaywalkers (high risk).

The Old Way: Current safety checks look at one single frame of video at a time. If the car predicts a path that looks "okay" for that split second, it assumes everything is fine.
The Reality: A car might make small, harmless mistakes for a few seconds, but then those mistakes pile up like a snowball rolling down a hill. By the time the snowball is huge (a crash), the old system is too slow to react.

2. The Insight: The "Mood Swing" of Errors

The authors noticed something interesting about how prediction errors behave. Even when the car is driving normally, its errors aren't random. They have "moods":

The "Calm" Mode: Traffic is flowing. The car predicts well. Errors are tiny.
The "Stressed" Mode: Traffic is heavy. The car is guessing. Errors get bigger.

Crucially, these modes switch back and forth like a light switch, but they follow a pattern (a Hidden Markov Model). The car doesn't just jump from "perfect" to "crashing"; it usually drifts from "calm" to "stressed" and stays there for a while. The old systems missed this pattern because they treated every second as an isolated event.

3. The Solution: The "Cumulative Scorecard"

The authors created a new monitoring system called DC-MMD. Think of it as a scorecard that the co-pilot keeps in their pocket.

The Baseline (The "Normal" Score): First, they teach the system what "normal" driving looks like using data from the car's training. They create a reference profile of what errors should look like when the car is doing well.
The Watchdog (The "CUSUM" Rule): As the car drives, the system doesn't just look at the current error. It looks at a stream of errors.
- If the car makes a small mistake, the scorecard goes up a tiny bit.
- If the car makes a few small mistakes in a row, the scorecard goes up more.
- The Magic: The system uses a mathematical trick (Maximum Mean Discrepancy) to compare the current stream of errors against the "Normal" profile. It doesn't need to know exactly what the new, dangerous situation is (e.g., it doesn't need to know it's snowing). It just needs to know that the errors look different from the normal pattern.
The Alarm: Once the scorecard crosses a specific line (a threshold), the system triggers an alarm: "OOD Event Detected!" (Out-of-Distribution). This tells the car's main computer, "Stop trusting the current prediction; switch to a safe backup plan immediately."

4. Why This is a Game-Changer

The paper highlights three superpowers of this new monitor:

It's Fast (Provable Guarantees):
Imagine a smoke detector. You want it to go off immediately when there's a fire, but you don't want it to go off every time you toast bread (false alarm). This system is mathematically proven to find the "fire" (the dangerous situation) as fast as physically possible without screaming "fire" every time you make toast.
It's Tough (Robust to Chaos):
Real-world data is messy. Sometimes errors are huge outliers (like a car suddenly swerving). Many systems break when faced with these "heavy-tailed" weird events. This system is like a shock absorber; it keeps working smoothly even when the data gets weird and unpredictable.
It Doesn't Need a Crystal Ball:
Most safety systems need to know exactly what the danger looks like beforehand (e.g., "If it's snowing, do X"). But in the real world, you can't predict every possible disaster. This system works like a detective who doesn't need to know the suspect's name to know a crime is happening. It just sees that the behavior is wrong compared to the norm.

The Bottom Line

This paper gives self-driving cars a new kind of "gut feeling." Instead of just blindly trusting their AI predictions, they now have a lightweight, mathematically rigorous watchdog that watches the pattern of mistakes. If the pattern changes, the watchdog sounds the alarm instantly, allowing the car to switch to a safe mode before a crash occurs.

It's the difference between a driver who only looks at the road ahead and a driver who also has a co-pilot constantly checking the dashboard, listening to the engine, and saying, "Wait, that sound isn't right, let's pull over now."

1. Problem Statement

In safety-critical Cyber-Physical Systems (CPS), particularly autonomous driving, trajectory prediction models often degrade in performance when encountering Out-of-Distribution (OOD) scenarios (e.g., rare weather, complex interactions, or unseen traffic patterns).

The Challenge: Existing Out-of-Distribution (OOD) detection methods, such as Uncertainty Quantification (UQ) (e.g., Bayesian Neural Networks, Ensembles), often fail in this context because:
1. They operate on a frame-by-frame basis, ignoring the temporal accumulation of small errors that lead to catastrophic failures.
2. They incur high computational latency, making them unsuitable for real-time multi-agent forecasting.
3. They rely on assumptions that training data represents operational conditions, which fails under severe distribution shifts.
The Goal: Develop a monitoring framework that detects OOD events in trajectory prediction errors with provable guarantees on detection delay and false-alarm rates, while remaining robust to unknown post-change distributions and heavy-tailed noise.

2. Methodology: DC-MMD

The authors propose DC-MMD (Data-driven Cumulative Maximum Mean Discrepancy), a framework that reframes OOD monitoring as a Quickest Change-Point Detection (QCD) problem driven by a Hidden Markov Model (HMM).

A. Modeling Prediction Errors as an HMM

The core insight is that prediction errors in autonomous driving are non-stationary even under normal (In-Distribution, ID) conditions.

Latent Modes: Errors alternate between a Low-Risk Mode (stable, routine traffic) and a High-Risk Mode (complex interactions, dense traffic).
HMM Structure: The error sequence $\{e_t\}$ ${e_{t}}$ is modeled as an HMM where:
- States: $Z_t \in \{L, H\}$ (Low-error, High-error).
- Transitions: Governed by a transition matrix $P$ (capturing the persistence of error modes).
- Emissions: Errors are drawn from state-specific distributions ( $h_L, h_H$ ).
Observation: The authors empirically validate that a 2-state Gaussian HMM accurately captures the dynamics of prediction errors across diverse datasets (NGSIM, nuScenes, ApolloScape).

B. The Detection Algorithm (DC-MMD)

Instead of relying on a companion reference sequence (which is unavailable in online deployment), DC-MMD constructs a reference distribution from the second-order statistics of the ID training data.

Second-Order Sampling: The algorithm monitors pairs of consecutive errors $x_t = (e_{t-1}, e_t)$ to capture temporal dependence.
Block Partitioning: The error stream is partitioned into non-overlapping blocks of length $m$ .
MMD Statistic: For each block, the algorithm computes the Maximum Mean Discrepancy (MMD) between the empirical distribution of the current block and the reference ID distribution ( $\lambda^{(0)}$ $λ^{(0)}$ ) in a Reproducing Kernel Hilbert Space (RKHS).
- MMD measures the distance between distributions without requiring parametric assumptions.
CUSUM Decision Rule: A cumulative sum statistic $W_t$ is updated iteratively:
$W_t = \max\{0, W_{t-1} + (D(\hat{\mu}_t, \lambda^{(0)}) - \zeta)\}$
Where $\zeta$ is an offset to control sensitivity to noise. An OOD event is declared when $W_t$ exceeds a threshold $b$ .

C. Theoretical Guarantees

The method provides rigorous statistical guarantees:

Detection Delay: The Worst-Case Average Detection Delay (WADD) is bounded and achieves order-optimality ( $O(\log \gamma)$ ), matching the theoretical lower bound for QCD problems.
False Alarm Control: The Mean Time to False Alarm (MTFA) grows exponentially with the threshold, allowing strict control over false positives.

3. Key Contributions

Latent-Aware Detection: The first framework to cast runtime OOD monitoring for trajectory prediction as a QCD problem over an HMM-driven error process, explicitly modeling the latent switching dynamics of traffic.
Robustness to Heavy-Tailed Distributions: Unlike likelihood-based methods (e.g., G-CUSUM) that assume Gaussian noise, DC-MMD uses kernel-based MMD, making it robust to heavy-tailed error distributions (e.g., Student-t, Laplace) common in real-world driving.
Distribution-Free Post-Change Handling: The method does not require prior knowledge of the post-change distribution ( $\Theta^{(1)}$ ). It only requires that the pre- and post-change distributions are distinguishable, making it suitable for unknown OOD scenarios.
Real-World Validation: Extensive evaluation on three major datasets (NGSIM, nuScenes, ApolloScape) and two distinct prediction architectures (GRIP++, FQA).

4. Experimental Results

The authors evaluated DC-MMD against state-of-the-art baselines, including Uncertainty Quantification methods (NLL, lGMM) and Sequential Change Detectors (G-CUSUM, GMM-CUSUM, Mode-CUSUM).

Detection Accuracy: DC-MMD achieved the highest AUROC across all datasets (e.g., 0.93 on NGSIM, 0.89 on nuScenes), significantly outperforming point-wise UQ methods (~0.6) and sequential baselines.
Delay vs. False Alarm: DC-MMD reduced the Worst-Case Average Detection Delay (WADD) by 15–35% compared to the strongest sequential baseline (Mode-CUSUM) while maintaining comparable Mean Time to False Alarm (MTFA).
Robustness to Heavy Tails: Under heavy-tailed noise (Student-t), likelihood-based methods saw delay increases of 30–52%, whereas DC-MMD's delay increased by only ~9%.
Unknown Post-Change: In scenarios where the post-change distribution was unknown, DC-MMD achieved detection delays comparable to "Robust CUSUM" (16.7 vs 16.4 steps) and significantly better than misspecified G-CUSUM (25.8 steps).
Computational Efficiency: The method operates with $O(1)$ complexity per step (via recursive sliding-window updates), making it suitable for real-time deployment with negligible overhead (< 0.01ms).

5. Significance

This work addresses a critical gap in autonomous driving safety: the lack of reliable, real-time monitoring for trajectory prediction models under distribution shifts.

Safety Assurance: By providing provable guarantees on detection delay and false alarms, DC-MMD enables downstream controllers (like MPC or Control Barrier Functions) to trigger safe fallback strategies before a prediction error leads to a collision.
Practical Deployment: The method is model-agnostic (works with any predictor), computationally lightweight, and robust to the non-stationary, heavy-tailed nature of real-world traffic data.
Theoretical Advancement: It bridges the gap between theoretical QCD (which often assumes known distributions) and practical deep learning monitoring by leveraging HMMs and kernel methods to handle latent dynamics and unknown shifts.