Technical Summary: CADET – Physics-Grounded Causal Auditing and Training-Free Deconfounding of End-to-End Driving Planners

1. Problem Statement

End-to-end (E2E) autonomous driving planners, trained via imitation learning on large logged datasets, are prone to causal confusion. These models often exploit spurious correlations—statistical shortcuts where scene elements (e.g., specific building facades, roadside objects) co-occur with expert actions in the training data but do not causally determine those actions. While such models may achieve low open-loop displacement errors (L2) and collision rates, their reliance on non-causal cues compromises reliability in long-tail or distribution-shift scenarios.

Current evaluation metrics (L2 displacement, collision rate) are dominated by ego status and fail to detect whether a planner relies on spurious cues. Furthermore, existing remedies, such as causal-intervention training (e.g., CausalVAD), require retraining large models and cannot audit or repair planners that are already deployed. There is a critical gap in training-free tools capable of auditing, benchmarking, and repairing spurious reliance in pretrained E2E planners without parameter updates.

2. Methodology: The CADET Framework

CADET (Causal Auditing and Deconfounding at Test-time) is a training-free framework designed to audit, benchmark, and repair spurious reliance in frozen E2E planners. It operates entirely via forward passes on a single GPU and comprises three core components:

A. Physics-Grounded Causal Reliance Score (PCR)

To distinguish true causal agents from spurious correlates, CADET fuses three signals into a per-query score ($PCR_i$):

1. Model Influence ($\tilde{I}_i$): Measures the change in the planned trajectory when a specific agent query is ablated (removed). High influence indicates the model relies on the agent.
2. Physics-Geometric Prior ($\rho_i$): An external, distribution-independent anchor derived from the perception module's outputs (3D box, velocity, class, confidence). It calculates the physical plausibility of an agent influencing the decision based on:
   • Time-to-Collision (TTC).
   • Drivable-corridor relevance (path overlap).
   • Dynamic vs. static state.
   • Detection confidence.
   • Function: Acts as a "hard gate." If an agent is physically plausible ($\rho_i$ is high), it is protected from being flagged as spurious, regardless of model influence.
3. Cross-Environment Stability ($\sigma_i$): Measures the variance of model influence across different environments (e.g., weather, time of day). High variance flags locally spurious reliance that the physics prior might miss.

The final score is computed as $PCR_i = \tilde{I}_i \cdot (1 - \rho_i) \cdot (1 + \lambda \sigma_i)$, subject to a hard gate that prevents flagging physically plausible agents.

B. Counterfactual Robustness Benchmark

CADET introduces a standardized protocol to quantify reliance on spurious vs. causal factors using three perturbation families and corresponding metrics:

• Spurious Perturbation: Removing physically irrelevant agents. Metric: Causal Stability Index (CSI) (measures plan stability; higher is better).
• Causal-Link Perturbation: Modifying causal agents (e.g., braking lead vehicle). Metric: Causal Response Index (CRI) (measures correct response direction; higher is better).
• Distribution-Shift Perturbation: Applying style shifts (e.g., day to night) while preserving causal structure. Metric: Causal Consistency Score (CCS) (measures plan consistency; higher is better).

C. Test-time Causal Masking (TCM)

At inference, CADET suppresses flagged spurious queries (those with high $PCR_i$ and low $\rho_i$) and re-plans using only the remaining causal agents. This acts as an approximate $do(\cdot)$ intervention, removing spurious confounding at deployment with negligible overhead and no training.

3. Key Contributions

1. Physics-Grounded Causal Audit (PCR): A training-free, model-agnostic score that identifies spurious reliance by fusing perturbation-based influence with a physics-geometric prior. It serves as a hard external arbiter independent of training data statistics.
2. Counterfactual Robustness Benchmark: A standardized protocol with three perturbation families and metrics (CSI, CRI, CCS) to quantify the extent of spurious reliance in pretrained planners.
3. Training-Free Deconfounding (TCM): A method to suppress flagged spurious queries at inference, improving counterfactual robustness without retraining or parameter updates.
4. Reproducibility: The entire pipeline is inference-only, runs on a single 16 GB GPU, and is released as an open-source toolkit.

4. Experimental Results

The authors evaluated CADET on two settings: a controlled synthetic benchmark (SpurGen) and a real-world audit of a pretrained SparseDrive planner on nuScenes.

• SpurGen (Synthetic): In a benchmark where spurious/causal status is known by construction, single-signal baselines failed:
  • Influence-only: High recall but low precision (0.35–0.39) because it cannot distinguish between causal agents and spurious shortcuts with high influence.
  • Physics-only: Low precision (<0.20) because it flags all physically implausible objects, including benign distractors the model ignores.
  • Invariance-only: Low precision and recall due to the "faithfulness assumption" failure (global spurious correlations appear invariant).
  • CADET (PCR): Achieved 0.942 precision and 0.886 recall (F1: 0.913), successfully isolating spurious dependencies by requiring both high influence and physical implausibility.
• nuScenes (Real-world Audit):
  • The pretrained SparseDrive planner was found to be predominantly causal (9.5× higher influence for causal agents vs. spurious ones).
  • However, in specific frames, the most influential physically irrelevant agent affected the plan as strongly as an average causal agent.
  • TCM Effect: Applying TCM removed spurious reliance, raising the Causal Stability Index (CSI) from ~0.47 to ~0.96.
  • L2 Blindness: Crucially, while TCM significantly altered the planner's causal behavior, the open-loop L2 displacement error remained virtually unchanged (1.90 $\to$ 1.98). This empirically demonstrates that standard displacement metrics are blind to causal robustness.

5. Significance and Claims

The paper claims that CADET addresses a critical gap in the deployment of E2E autonomous systems: the inability to audit and repair causal confusion in frozen, already-deployed models.

• Methodological Shift: The authors argue that open-loop L2 displacement is insufficient for evaluating E2E planners because it is dominated by ego status and insensitive to whether the planner reasons over causal or spurious cues. They propose that causal-robustness metrics (like CSI and per-agent influence statistics) must accompany displacement error.
• Practical Utility: Unlike prior causal methods that require retraining, CADET offers an immediate, training-free diagnostic and repair mechanism. It allows developers to identify "hidden" reliance on spurious features (e.g., a specific billboard or mailbox) that could lead to failure in long-tail scenarios.
• Limitations: The authors modestly note that CADET relies on the quality of the upstream perception module; if the detector fails, the physics prior is compromised. Additionally, while it improves reliability via a physics anchor, it does not provide a mathematical guarantee of separating true causes from globally spurious correlates under all conditions of faithfulness failure.

In summary, CADET provides a necessary tool for the "audit and repair" phase of E2E driving, demonstrating that strong planners can still harbor severe, localized causal confusions that standard metrics fail to detect.