acceptodds
Withdrawn from ICLR 2027

CausalDrive: Counterfactual Causal Decomposition for End-to-End Autonomous Driving

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

End-to-end autonomous driving typically learns driving policies by imitating expert trajectories. However, observed expert trajectories inherently couple nominal driving intentions with multimodal responses to surrounding agents. Different combinations of these factors can produce similar trajectories, resulting in severe causal ambiguity that makes it difficult for data-driven methods to identify their respective contributions to the final trajectory. To address this ambiguity, we propose CausalDrive, a causal decomposition framework that separates driving policy learning into two sequential stages: deterministic nominal intention prediction and probabilistic interactive response modeling. Specifically, we first construct a counterfactual generator that produces nominal trajectories from navigation commands and map topology, providing independent supervision for our desighed Nominal Intention Planner. We then develop a diffusion-based Interactive Response Policy to model the multimodal residual between nominal and expert trajectories, thereby capturing the uncertainty of interaction-induced behavior. The outputs of the two stages are composed to produce the final interaction-aware trajectory. Extensive experiments demonstrate that CausalDrive improves overall trajectory quality and exhibits significant advantages in complex interactive driving scenarios.

Primary area:
applications to robotics, autonomy, planning
Keywords:
End-to-End Autonomous Driving, Causal Inference

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_kkMUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CausalDrive, an autonomous driving framework consisting of two major components: (1) a Nominal Intention Planner (NIP) and (2) a Style Execution Policy (SEP). Experimental results demonstrate strong benchmark performance, achieving good performance on NAVSIM v1 and NAVSIM v2.

Critical strengths
  1. NIP + SEP seems to be novel
  2. Causal ambiguity is critical in driving domains
Critical weaknesses
  1. Limited real-world validation
  2. Incomplete related work
  3. Limited methodological novelty
Questions for authors

CausalDrive achieves 91.2 PDMS on NAVSIM v1 and 89.5 EPDMS on NAVSIM v2

  1. Although the proposed method achieves strong benchmark performance, these results alone do not sufficiently establish its safety, robustness, or generalizability in real-world driving environments. While the authors provide additional evaluations on the NAVSIM navhard subset, the experiments remain limited to benchmark-based assessments. Could the authors provide more comprehensive evaluations in realistic, reactive closed-loop environments, particularly under safety-critical scenarios, unforeseen traffic interactions, and out-of-distribution conditions?

  2. Incomplete discussion of related work. The related-work section overlooks several relevant studies on causal imitation learning, counterfactual reasoning, and generative modeling for autonomous driving, including [1][2][3]

CausalDrive, a causal decomposition framework that models nominal intention deterministically and driving style probabilistically

  1. Limited methodological novelty. The proposed framework appears to be an incremental extension of existing diffusion-based planning approaches, particularly DiffusionDrive and DiffusionDriveV2. Although the ablation studies demonstrate improvements from these components, the results do not clearly establish whether the gains arise specifically from causal reasoning or from the additional supervision and architectural decomposition.

    3.1 Could the authors provide a more rigorous justification for why causal reasoning is necessary for the proposed framework? Additional comparisons with non-causal trajectory decomposition or residual-learning baselines would help isolate the contribution of the claimed causal mechanism.

    3.2 Given that ResAD [4] already introduces residual trajectory modeling, could the authors clarify the fundamental methodological differences between these approaches and the proposed framework?

    3.3 It remains unclear why the proposed decomposition necessarily requires causal reasoning rather than conventional conditional modeling. Could the authors specfically explain what aspects of the framework are uniquely enabled by causality and cannot be achieved through standard trajectory decomposition or residual learning?

    3.4 Insufficient theoretical justification of the causal framework. The causal formulation requires further clarification. Section 3.1 introduces a causal graph and uses the intervention notation do to motivate counterfactual nominal trajectory generation. However, the underlying causal assumptions and their relationship to a formal structural causal model (SCM) are not sufficiently established.

4 Lack of statistical reliability

4.1 Evaluation error bars? Could the authors report results across multiple random seeds, including standard deviations or confidence intervals, to demonstrate the statistical reliability of the reported performance improvements?

4.2 Missing ablations about causal mechanism: Could the authors provide ablations comparing the proposed causal decomposition with non-causal residual-learning baselines under identical architectures and training settings to demonstrate that the improvements arise specifically from causal reasoning?

Lots of other ablations on matched architectures vs. non-causal, training objectives, and candidate-ranking procedures seem to be missing.

  • Typos?

We then develop a diffusion-based \textbf{Interactive Response Policy} to model the multimodal residual between vs. We then develop a diffusion-based Style Execution Policy to capture the multimodal residual between nominal

[1] "Learning human driving behaviors with sequential causal imitation learning." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 36. No. 4. 2022.

[2] "CausalDrive: Real-time causal world models for autonomous driving." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.

[3] "Causalaf: Causal autoregressive flow for safety-critical driving scenario generation." Conference on robot learning. PMLR, 2023.

[4] "ResAD: Normalized residual trajectory modeling for end-to-end autonomous driving." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

Reviewer_pj55OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CausalDrive, an end-to-end driving framework that decomposes trajectory generation into nominal intention prediction and multimodal residual modeling. A rule-based generator constructs nominal trajectory targets from navigation, map topology, and ego state. The Nominal Intention Planner (NIP) learns these targets, while a diffusion-based Style Execution Policy (SEP) generates residual adjustments conditioned on the nominal trajectory and scene context. A trajectory ranker selects the final plan. Experiments on NAVSIM v1, v2, and challenging interactive scenarios report competitive planning performance.

Critical strengths

1.The framework provides a clear structural separation between route-consistent nominal planning and interaction-dependent adjustments. The independently constructed nominal targets offer an intuitive inductive bias.

2.Combining deterministic nominal prediction with probabilistic residual generation is a reasonable approach to representing multiple possible executions of a driving intention.

3.The evaluation covers two NAVSIM versions and a challenging scenario split. Component-level metrics and ablations help illustrate the complementary roles of nominal planning, residual generation, and trajectory ranking.

Critical weaknesses

1.Existing planners already receive navigation, map, ego-state, and surrounding-agent information, which may allow them to distinguish route requirements from interaction responses implicitly. The examples motivate the proposed decomposition, but do not establish when this ambiguity causes planning errors or how much the explicit separation resolves them.

2.A scene may admit multiple reasonable nominal trajectories, so using a single rule-generated target may be restrictive. The paper should examine sensitivity to alternative valid references and whether SEP can correct inaccurate NIP predictions.

3.Section 3.3 states that rule-generated nominal trajectories replace NIP predictions as SEP conditions with a probability of 10%. However, Equation (7) defines the residual target relative to the NIP prediction. It is unclear whether the residual target and reconstruction baseline are updated consistently when the condition is replaced.

4.Table 4 reports a PDMS@1 of 95.5 for CausalDrive, whereas Table 1 reports a final PDMS of 91.2. The paper should explain whether these results use different scene subsets, ranking procedures, or aggregation rules.

Questions for authors

1.How does the rule generator choose among multiple reasonable nominal trajectories? Have the authors examined sensitivity to alternative lane references or speed profiles?

2.In the 10% replacement cases, is the residual target recomputed relative to the rule-generated trajectory? Which nominal trajectory is used to reconstruct the complete trajectory, and why is this replacement expected to reduce exposure bias?

3.Is any analysis available showing how final planning performance changes with NIP prediction error? Can SEP correct substantial nominal errors while preserving navigation consistency?

Reviewer_LjMxOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a trajectory planning framework for autonomous driving comprising two modules. The nominal intention planner is a regression model trained with an L1 loss against rule-based nominal trajectories. The style execution policy uses a diffusion model to generate residual corrections between the initial prediction and the ground-truth expert trajectory. Its stochastic sampling process produces multiple candidate trajectories, which are evaluated by a learned ranker trained on simulator-derived scores.

Critical strengths
  1. The paper is clearly written, and the overall framework is easy to understand.
  2. The method is conceptually simple.
Critical weaknesses

1. Limited novelty

Initial prediction followed by refinement is a well-established strategy in trajectory prediction. Within vehicle trajectory planning and prediction, related approaches include DeMo [1], FiM [2], and GoIRL [3]. The manuscript does not clearly articulate what distinguishes its proposed decomposition from these approaches.

2. Insufficient detail for reproducibility

The procedure for constructing nominal trajectories is described only at a high level. Although the paper outlines components involving path planning, geometry, and kinematics, it does not provide sufficiently concrete algorithms or implementation details to reproduce the resulting training targets. Because this construction is presented as a central contribution and directly determines the supervision for the nominal intention planner, a more complete specification is necessary.

3. Uncertainty in the reported performance differences

Several leading methods achieve closely clustered scores on the NAVSIM benchmarks reported in Tables 1 and 2. Without estimates of variability, it is difficult to determine whether the reported differences reflect consistent improvements or variation arising from training and stochastic trajectory sampling. The authors should report means and standard deviations across multiple independent runs and specify the number of runs. Comparisons should use consistent evaluation protocols and sampling budgets.

[1] Zhang et al., “DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States,” NeurIPS 2024.

[2] Pei et al., “Foresight in Motion: Reinforcing Trajectory Prediction with Reward Heuristics,” ICCV 2025.

[3] Pei et al., “GoIRL: Graph-Oriented Inverse Reinforcement Learning for Multimodal Trajectory Prediction,” ICML 2025.

Questions for authors
  1. What is the contribution beyond existing prediction-and-refinement frameworks?
  2. How exactly are the nominal training trajectories constructed? For example, can authors provide a reproducible algorithm, including lane selection and/or geometric path construction?
  3. Are the reported improvements consistent across independent runs?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.