🤖 machine learning

Can Attribution Predict Risk? From Multi-View Attribution to Planning Risk Signals in End-to-End Autonomous Driving

This paper proposes a hierarchical attribution framework for end-to-end autonomous driving that extracts statistical signals from multi-view inputs to effectively predict planning risks and identify potential collisions without relying on auxiliary monitoring models.

Original authors: Le Yang, Ruoyu Chen, Haijun Liu, Jiawei Liang, ShangQuan Sun, Xiaochun Cao

Published 2026-05-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Le Yang, Ruoyu Chen, Haijun Liu, Jiawei Liang, ShangQuan Sun, Xiaochun Cao

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 robot to drive a car. In the old days, we built the robot's brain in separate rooms: one room for "seeing," another for "thinking," and a third for "steering." But recently, engineers have built "End-to-End" robots. These are like a single, super-smart brain that looks at the road and immediately decides where to drive, skipping all the middle steps.

The problem? These super-brains are black boxes. They make decisions, but we don't know why. If the robot suddenly swerves into a tree, we can't easily tell if it was confused by a shadow, a weird sign, or a glitch.

This paper asks a big question: Can we peek inside the robot's brain to see what it's looking at, and use that to predict if it's about to make a dangerous mistake?

Here is how the authors solved this, explained with simple analogies:

1. The "Six-Eyed" Detective

Most self-driving cars have six cameras (like a human with six eyes looking in all directions). The robot takes all six views and decides on a path.

The Old Way: Previous methods tried to explain the robot's mind by writing a text summary (like a diary entry) or by training a separate "guard dog" to watch the robot. But text can be vague, and the guard dog doesn't know exactly what the robot is thinking right now.
The New Way: The authors created a tool called Hierarchical Attribution. Think of this as a high-tech flashlight that the robot uses to scan its own six camera feeds. It highlights the specific pixels (tiny dots of the image) that the robot relied on to make its decision.

2. The "Coarse-to-Fine" Search

Scanning every single pixel in six cameras is too slow and confusing. So, the authors designed a smart search strategy:

Step 1 (Coarse): Imagine looking at a map of a city. First, you just look at the neighborhoods (e.g., "The intersection ahead" or "The car to the left"). The robot quickly ranks which neighborhoods are most important.
Step 2 (Fine): Once the robot knows the important neighborhood, it zooms in to look at the specific houses and streets within that area.
This allows the system to find the exact visual clues the robot used without getting overwhelmed.

3. The Three "Risk Signals"

Once the robot highlights the important parts of the image, the authors didn't just look at the picture; they turned the highlights into three simple numbers (statistics) to act as a "risk alarm."

Think of these numbers as checking if the robot is too focused or too scattered:

Signal 1: The "Tunnel Vision" Meter (Attribution Entropy)
- The Metaphor: Imagine a detective solving a crime. A good detective looks at many clues (footprints, fingerprints, witness statements). A bad, risky detective might only stare at one clue (a single muddy shoe).
- The Math: If the robot's attention is spread out over many parts of the image, the number is high (Safe). If it's staring intensely at just one tiny spot, the number is low (Risky).
Signal 2: The "Cluster" Meter (Spatial Variance)
- The Metaphor: Imagine a student taking a test. A good student spreads their focus across the whole page. A risky student might focus only on the top-left corner of the page and ignore the rest.
- The Math: This checks if the robot's attention is clumped together in one small corner of a single camera view. If it is, that's a sign of risky "clumping."
Signal 3: The "One-Eye" Meter (Cross-Camera Gini)
- The Metaphor: You have six eyes. If you are driving safely, you use all of them. If you are driving dangerously, you might be ignoring your left and right eyes and only staring through your front windshield.
- The Math: This checks if the robot is ignoring 5 of its 6 cameras and relying too heavily on just one. If the attention is unbalanced, the number goes up (Risky).

4. The Results: Does it Work?

The team tested this on three different advanced robot drivers (BridgeAD, UniAD, and GenAD) using a massive dataset of real driving videos.

The Prediction: They found that when these three "Risk Signals" went up (meaning the robot was too focused, too clumped, or too one-sided), the robot was much more likely to make a mistake (like missing a turn or almost hitting a car).
The Accuracy: Their system could predict a potential crash about 77% of the time (a score known as AUROC). This is significantly better than random guessing.
The Generalization: Even when they tested the system on new scenes it had never seen before, the risk signals still worked. It wasn't just memorizing the old videos; it was actually understanding the robot's behavior.

5. Why This Matters

The authors aren't saying this system is a "fix" that stops crashes. They are saying it's a diagnostic tool.

Just like a doctor uses a thermometer to see if a patient has a fever (a signal that something is wrong, even if the thermometer doesn't cure the flu), this system uses "attribution signals" to tell engineers: "Hey, this robot is relying on a weird, narrow set of visual clues. It's about to make a mistake."

This allows developers to find dangerous scenarios faster and understand why their robots are failing, making self-driving cars safer and more transparent.

Technical Summary: Attribution Predicts Risk in End-to-End Autonomous Driving

Problem Statement
End-to-end autonomous driving models map raw multi-view sensor inputs directly to planning trajectories, offering improved system integration but introducing opaque decision-making processes. While existing interpretability methods generate textual explanations or train auxiliary monitoring models, they often fail to reveal the specific visual evidence underlying trajectory generation or remain decoupled from the planning process. A critical gap exists in determining whether attribution maps—typically used for post-hoc explanation—can serve as predictive signals for planning risk. Specifically, it is unclear if the distribution of a planner's reliance on visual evidence across six camera views correlates with trajectory errors or collision risks.

Methodology
The authors propose a hierarchical attribution framework designed specifically for multi-view end-to-end planners. The methodology consists of three main components:

Hierarchical Coarse-to-Fine Attribution:
- Input & Partitioning: The system takes synchronized six-view camera images and partitions each view into subregions using SLICO superpixel segmentation.
- Objective Function: Unlike classification-based attribution that maximizes class confidence, this framework uses L2 trajectory consistency as the objective. It defines a set function $F(S)$ based on sufficiency (keeping regions $S$ maintains the original trajectory) and necessity (removing regions $S$ causes significant deviation).
- Search Strategy: To address the computational cost of evaluating subsets across a six-view space, a coarse-to-fine greedy search is employed.
  - Coarse Stage: Adjacent subregions are merged into groups. The algorithm globally ranks these groups to identify high-contribution clusters.
  - Refinement Stage: Within the top-ranked groups, the algorithm performs a finer search over individual subregions, using the preceding groups as a conditional prefix. This produces a faithful per-pixel saliency tensor $T$ .
Derivation of Risk Signals:
Treating the saliency tensor as a probability distribution over the joint multi-camera pixel space, the authors extract three complementary statistics to quantify "over-reliance," hypothesizing that robust planning should avoid concentrating reliance on limited evidence:
- Attribution Entropy ( $H$ ): Measures the global concentration of attribution across the entire six-view space. Low entropy indicates reliance on a small set of visual locations.
- Within-Camera Spatial Variance ( $\sigma^2_{sp}$ ): Characterizes the spatial spread of attribution within individual camera views. Low variance indicates highly localized reliance within a view.
- Cross-Camera Gini Coefficient ( $Gini_{cam}$ ): Quantifies the imbalance of attribution mass across the six cameras. High values indicate disproportionate reliance on specific views.
Risk Prediction Framework:
These statistics are used as features to predict two planning risk proxies:
- Trajectory Error: Measured by Average Displacement Error (ADE).
- Collision Risk: A binary indicator of whether the predicted trajectory intersects an obstacle.
  The authors fit ridge regression for ADE and logistic regression for collision detection, evaluating performance via Spearman correlation and AUROC.

Key Contributions

Predictive Attribution: The paper investigates and validates that attribution can extend beyond post-hoc explanation to serve as a predictive signal for planning risk in end-to-end autonomous driving.
Multi-View Attribution Algorithm: A novel coarse-to-fine attribution method that treats six camera views as a unified search space, utilizing L2 trajectory consistency rather than classification confidence.
Risk Statistics: The definition of three statistics (Entropy, Spatial Variance, Gini) that quantify global, within-view, and cross-view over-reliance, respectively.
Empirical Validation: Comprehensive experiments on the nuScenes dataset using three distinct planners (BridgeAD, UniAD, GenAD) demonstrating the correlation between attribution patterns and risk.

Experimental Results
Experiments were conducted on the nuScenes validation set with BridgeAD, UniAD, and GenAD:

Risk Correlation: The attribution statistics show significant correlation with planning risk. For collision prediction, the joint model achieved an AUROC of 0.77 ± 0.04 across the three planners. For trajectory error, the Spearman correlation reached 0.30 ± 0.07.
Generalization: The risk signal generalizes to held-out scenes (80/20 split) with negligible degradation (AUROC 0.77 ± 0.09), indicating the signal is tied to attribution structure rather than specific scene object layouts.
Robustness: Replacing the proposed hierarchical attribution with a random-sampling baseline (RISE) yielded weaker but consistent signals (AUROC 0.67 ± 0.11), suggesting the risk statistics are not artifacts of the specific attribution algorithm.
Control Variables: The association between attribution and risk persists even when controlling for scene object configuration (e.g., object count, spatial spread), confirming the signal reflects the planner's visual dependency rather than mere scene complexity.
Efficiency: The hierarchical search is significantly faster than exact greedy search (EAGLE) while maintaining high faithfulness (closer to EAGLE than RISE).

Significance and Claims
The paper claims that attribution patterns provide useful warning signals for decision-level risk in end-to-end driving planners. By treating attribution as a measurement object rather than just a visualization tool, the framework enables scalable pre-deployment auditing. It allows developers to identify scenes where a planner's reliance on visual evidence is concentrated, spatially diffuse, or unevenly distributed in ways that co-occur with elevated risk.

The authors maintain a modest stance, emphasizing that these statistics are diagnostic signals rather than safety certifications. They note that while the signals correlate with risk at a population level, they do not guarantee per-sample safety. The framework is presented as a complementary tool to existing output-side safety evaluations (like trajectory error checks) and closed-loop simulation, helping to locate failure-prone scenarios more efficiently than visual inspection alone. The primary limitation acknowledged is the computational cost of high-fidelity attribution, which currently precludes real-time online monitoring, though the results suggest the potential for future real-time, high-fidelity attribution methods.