CausalDrive: Causal Decomposition for Data-Efficient End-to-End Autonomous Driving

Xiaobao Wei, Jingxian Wu, Lirong Yang, Qing Han, Yuan Gao, Ziwei Yang, Shuai Yang, Weiwei Liu, Chang HuangarXiv 2026 · arXiv

This summary provides a technical overview of the paper "CausalDrive: Causal Decomposition for Data-Efficient End-to-End Autonomous Driving."

TL;DR

CausalDrive is a novel end-to-end autonomous driving framework that decomposes driving behavior into two causally distinct components: a deterministic "nominal intention" (where to go) and a probabilistic "driving style" (how to interact). By separating these factors, it achieves state-of-the-art performance on major benchmarks while significantly improving data efficiency and robustness in complex interactive scenarios.

What Problem Does This Paper Address?

Most current end-to-end autonomous driving systems rely on imitation learning, where a model is trained to directly replicate expert trajectories. The authors argue that this approach suffers from causal ambiguity. An expert's path is an "entangled" result of their high-level intent (e.g., following a route) and their low-level execution style (e.g., yielding to a truck).

As illustrated in Figure 1, different combinations of intent and style can produce nearly identical trajectories. For instance, a car turning left might look very similar to a car going straight but swerving aggressively to avoid an obstacle. Forcing a single neural network to learn both from one target makes the learning process inefficient and confusing, especially when data is limited.

Figure 1: Causal ambiguity in direct imitation.

Methods

CausalDrive explicitly separates these factors using a two-stage architecture:

  1. Nominal Intention Planner (NIP): This module predicts a "nominal trajectory"—the path the car would take if the road were empty. It is trained using "counterfactual" targets derived solely from navigation commands, map topology, and the car's current state, ignoring other agents.
  2. Style Execution Policy (SEP): Conditioned on the nominal trajectory and the surrounding environment (other cars, pedestrians), this diffusion-based model predicts a "style residual." This residual represents the necessary adjustments to the nominal path to account for interactions.
  3. Trajectory Ranker: Since the SEP is probabilistic and can generate multiple plausible responses (yielding vs. overtaking), a coarse-to-fine ranker selects the final executable trajectory from the generated candidates.

This decomposition is grounded in a causal modeling assumption: navigation and maps determine intention, while surrounding agents influence execution.

Figure 3: The CausalDrive architecture overview.

Datasets

The authors evaluated CausalDrive on two major planning-centric benchmarks:

  • NAVSIM v1: A benchmark built on the nuPlan dataset, focusing on closed-loop metrics like PDMS (Planning-centric Driving Metric Score).
  • NAVSIM v2: A more challenging version with updated evaluation protocols (EPDMS). Training was performed on the navtrain subset (1,192 scenes) and evaluated on the navtest split.

Key Insights and Ideas

  • Reducing Redundancy: In routine driving, the nominal trajectory is often very close to the expert's actual path. By modeling residuals, the SEP can focus its learning capacity on the differences that matter—the complex interactions—rather than relearning basic lane-following.
  • Residual-Guided Data Mining: The size of the residual (the gap between nominal and expert paths) serves as a proxy for scenario difficulty. The authors found that prioritizing training on "high-residual" scenes (those with significant interaction) leads to much better performance than random sampling.
  • Structured Multimodality: By anchoring the diffusion process to a nominal trajectory, the model produces more structured and separable driving styles (e.g., distinct clusters for "yielding" vs. "following") compared to models that generate full trajectories from scratch.

Key Results

CausalDrive sets new state-of-the-art records for ResNet-34 based models:

  • Top Performance: It achieved 91.2 PDMS on NAVSIM v1 and 89.5 EPDMS on NAVSIM v2, matching or exceeding much larger models.
  • Superior Data Efficiency: With only 20% of the training data, CausalDrive outperformed baseline models trained on significantly more data. Figure 5 shows the clear gap in performance at lower data scales.
  • Interactive Excellence: In challenging scenarios where expert behavior deviates significantly from the nominal path (e.g., dense traffic), CausalDrive showed a massive +9.9 EPDMS improvement over the previous best method, DiffusionDriveV2.
Figure 5: Data efficiency comparison.

Limitations and Future Work

While CausalDrive demonstrates impressive efficiency, the paper notes a few areas for further exploration:

  • Nominal Path Quality: The system's performance is partially dependent on the quality of the rule-based nominal trajectory construction. Errors in map data or complex topological branches could impact the NIP's reliability.
  • Foundation Model Integration: The authors suggest that this causal decomposition framework could be integrated into larger, multi-modal foundation models for even greater generalization.
  • Interactive Scenarios: While it excels in current benchmarks, testing in even more extreme, long-tail interactive scenarios remains a goal for future work.

Summary updated

Abstract

Efficient use of expert demonstrations is essential for end-to-end autonomous driving. However, directly imitating expert trajectories couples nominal intention, which specifies where the vehicle should go, with multimodal driving styles, which determine how that intention is executed. Different combinations of these factors can yield similar expert trajectories, obscuring their respective causal contributions and forcing the policy to learn both from a single entangled target. This ambiguity limits how effectively the policy can exploit expert demonstrations. We propose CausalDrive, a causal decomposition framework for data-efficient driving policy learning that separates deterministic nominal intention prediction from probabilistic driving style modeling. Specifically, we construct counterfactual nominal trajectories from navigation command, map topology, and ego state by removing style dependent effects, providing a Nominal Intention Planner with supervision independent of expert trajectories. Conditioned on the predicted nominal intention and surrounding agents, a diffusion-based Style Execution Policy models diverse driving styles as a multimodal distribution over residuals between nominal and expert trajectories. This causally structured supervision reduces redundant learning of shared nominal behavior in routine scenarios, directing expert supervision toward informative execution variations and improving data utilization. CausalDrive achieves 91.2 PDMS on NAVSIM v1 and 89.5 EPDMS on NAVSIM v2, matching or surpassing state-of-the-art performance. Further experiments demonstrate strong performance with limited expert data and clear advantages in challenging interactive scenarios.
BibTeX
@misc{wei2026causaldrive,
  author = {Xiaobao Wei and Jingxian Wu and Lirong Yang and Qing Han and Yuan Gao and Ziwei Yang and Shuai Yang and Weiwei Liu and Chang Huang},
  title = {{CausalDrive: Causal Decomposition for Data-Efficient End-to-End Autonomous Driving}},
  year = {2026},
  eprint = {2611.24080},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.24080},
}
Figure