acceptodds
Under review as a conference paper at ICLR 2027

Geometrically Aligned Diffusion for End-to-End Autonomous Driving

Abstract

While generative diffusion models have emerged as a powerful paradigm for multimodal trajectory generation, their deployment in safety-critical autonomous driving is fundamentally bottlenecked by a critical misalignment: unconstrained statistical generation frequently violates physical, geometric, and kinematic constraints. To address this, we propose STEER, a unified framework that systematically embeds physical inductive biases into the generative lifecycle through a two-fold alignment strategy. At the macro-geometric level, STEER grounds the generative manifold in physical space by introducing Scene-Conditioned Anchor generation (SCA) for topology-compliant prior initialization, while leveraging Graph Attention Agent Interaction (GAI) to enforce spatial locality and prevent attention entropy collapse during iterative denoising. At the micro-kinematic level, STEER guarantees high-order trajectory feasibility and deterministic execution by coupling Adaptive Dual-Path Refinement (ADPT) with a Hierarchical Trajectory Selection (HTS) mechanism. Specifically, ADPT fuses global intent with local autoregressive modeling for kinematic smoothness, while HTS decodes the generative density through joint optimization over predictive utility, topological boundaries, and motion feasibility. Extensive experiments on the NAVSIM benchmark demonstrate that STEER establishes a new state-of-the-art, successfully bridging statistical generative modeling with physically grounded embodied planning.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Generative diffusion models, Autonomous driving trajectory generation, Physical constraint alignment, Embodied motion planning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_RcSHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents STEER, a diffusion-based end-to-end driving framework designed to improve the geometric and kinematic feasibility of generated trajectories. STEER consists of four main components:

  1. Scene-Conditioned Anchor generation (SCA), which adapts trajectory anchors to the current scene;
  2. Graph Attention-based Agent Interaction (GAI), which introduces distance-based locality into agent interaction;
  3. Adaptive Dual-Path Trajectory Refinement (ADPT), which combines one-shot global prediction with segment-wise autoregressive refinement;
  4. Hierarchical Trajectory Selection (HTS), which applies learned scoring, kinematic penalties, hard constraints, and PDM-based reranking. The paper reports a PDMS of 93.8 on NAVSIM-v1 and an EPDMS of 87.2 on the reported NAVSIM-v2 navtest protocol. And it operates at 128.36 ms per frame, corresponding to 7.79 FPS on an NVIDIA A100.
Critical strengths
  1. The paper addresses a practical limitation of diffusion planners.、
  2. This paper is easy to understand and follow.
  3. The paper reports useful performance-latency ablations.
Critical weaknesses
  1. I find the paper to be a reasonable engineering integration of several physically motivated techniques such as map informations and hard constraints. I am not convinced that it constitutes a sufficiently novel or rigorously validated contribution for ICLR. In this context, describing STEER as “a next-generation diffusion framework” seems overstated.
  2. compared with recent strong baselines such as DrivoR (CVPR 2026), STEER does not show a clear advantage. The experimental results on NAVSIM are also not sufficiently compelling to support the claimed state-of-the-art performance.
  3. The paper lacks generalization and reactive closed-loop validation in stronger simulators such as AlphaSim or HUGSIM.
Questions for authors
  1. Is STEER camera-only or camera-LiDAR? What exactly is meant by the “shared image-LiDAR backbone”?
  2. Can the authors report multiple runs with standard deviations and a leave-one-out ablation starting from the full system?
  3. Does the 128.36 ms latency include perception and all selection stages?
  4. How much of the improvement remains when all methods use the same number of candidate trajectories and the same PDM reranking budget?
Reviewer_S1nkOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes STEER, a diffusion-based end-to-end autonomous driving framework designed to improve physical feasibility of generated trajectories. 4 core components are introduced to enhance the planning performance. However, the empirical results do not provide enough support for the efficacy of the modules and the design motivation.

Critical strengths

The paper tries to address a relevant problem in autonomous driving, proposing a technically feasible framework. 2. SCA is proven to be effective, providing the denoising model with a smaller set of trajectories for the current scene.

Critical weaknesses
  1. Factual error/conflicting claims exist in the paper: The description of EC in navsim v2 metrics is inaccurate (sec 4.1.2). EC penalizes the inconsistency between consecutive frames instead of the jerk and lateral acceleration in a single trajectory (which the ADPT/HTS try to address). Furthermore, although EC is highlighted, it is dropped in the later Table 2. From the presented scores, its HC is the lowest among all compared models, contradicting the original claim and the quantitative results shown in the appendix.

  2. Unsupported claim: In Sec 3.3 and Sec 4.3, ADPT is said to be optimized for both low-frequency (global) and high-frequency (local) output. While the comfort score is saturated in Table 4, the smoothness of trajectories can not be represented by EP or safety metrics. The ablation can only show that the model is more aggressive with dual-path fusion. To support this claim, an ablation on unsaturated EPDMS HC would help.

  3. Vague description of the module input. Especially the 3D BEV semantic map used in SCA and HTS.

Questions for authors

Please clarify the input of the GAI and the full HTS. Does the PDM reranking use privileged information as input, or does it use the model output as the scorer input? The current description in Sec 3.4 is vague. If the "high-resolution BEV" here is from the model itself, do you have a quantitative score for the perception quality?

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

The paper injects physical inductive biases at four points of a diffusion planning pipeline: scene-conditioned anchors (SCA) produced by a zero-initialized offset MLP, distance-biased graph attention for agent interaction (GAI), adaptive dual-path refinement that fuses one-shot and segment-wise autoregressive predictions (ADPT), and hierarchical trajectory selection (HTS) that ends in PDM reranking. With a ResNet-34 backbone it reports 93.8 PDMS on NAVSIM v1 and 87.2 EPDMS on NAVSIM v2 navtest and claims state-of-the-art on both leaderboards. Ablations cover the four modules, the dual-path design, the selection stages, the initialization strategy, the number of denoising steps and the number of modes.

Critical strengths
  • Scene-conditioned initialization: I appreciate the idea of conditioning the diffusion starting manifold on the scene instead of on offline clustering, and the zero-initialized offset MLP that makes SCA an identity map at the start of training is a clean design.
  • Test-time search made visible: Table 5 separates the learned scorer, the kinematic and hard-constraint filters and the PDM reranking, which is the right ablation to show how much comes from test-time search.
  • Efficiency: Latency and FPS on GPU (A100) are reported for every stage of the ablation (Table 3).
Critical weaknesses
  • The state-of-the-art claim does not hold. Table 1 is not a leaderboard snapshot but SparseDriveV2’s Table 2, reproduced with identical values and identical row ordering for all 17 ResNet-34 entries (its two V2-99 rows are dropped), so leaderboard entries scoring above 93.8 PDMS are absent. By May 2026 the public navtest leaderboard already listed several, e.g. ChainFlow-VLA (94.85), DrivoR (94.59) and CLOVER (94.49). At most, the claim holds among ResNet-34 methods, and it should be stated that way.
  • Mismatched EPDMS protocol on v2 (Table 2). The EPDMS values quoted for the strongest baselines, SparseDriveV2 at 86.7 and DiffusionDriveV2 at 85.5, are the figures from the original NAVSIM v2 code, which SparseDriveV2 reports separately (EPDMS∗) for fair comparison with prior work; under the updated official implementation SparseDriveV2 reports 90.1 and 87.5 (https://arxiv.org/pdf/2603.29163), both above STEER's 87.2. The paper does not say which code produced 87.2: if it is the updated one, STEER is third on v2 rather than first; if it is the original one, the comparison is matched but rests on a superseded metric. Finally, the EC column is present in both source tables but absent from Table 2, although Sec. 4.1.2 introduces EC as precisely the measurement that makes v2 relevant to this contribution
  • Test-time selection by the evaluation metric (Sec. 3.4). HTS selects “the trajectory with the highest PDM score”, i.e. by the metric that is then reported, but the paper never states which agent futures, map and drivable area this scorer uses at inference. If they come from the navtest logs, the +1.0 PDMS of the reranking step (92.8 to 93.8, Table 5) and the record NC and DAC are leakage rather than planning quality. The reranker inputs must be specified, and Tables 1 and 2 should also report results without reranking.
  • Inconsistent ablations. Credit for the central module is inconsistent, with Table 3 attributing +0.7 PDMS to SCA (89.2 to 89.9) and Table 6 attributing +3.3 to the same module (90.5 to 93.8), which cannot be reconciled from what the paper reports. In addition, the one-shot-only variant of ADPT (90.4, Table 4) scores below the model without ADPT (90.7, Table 3), and the default is 10 anchors in Sec. 4.1.3 but 20 modes in Appendix B.0.2.
  • The ablation is also unanchored because the “vanilla diffusion baseline” is never identified, so the reader cannot tell what the four modules are being credited against. The term “vanilla diffusion baseline” comes from DiffusionDrive (Liao et al., 2025) / DiffusionDriveV2. The implementation details in Sec. 4.1.3 coincide with DiffusionDrive's released configuration, but the paper does not state that it builds on it.
  • No variance estimates. No seeds, repeated runs or variance estimates appear anywhere, yet differences of 0.1 to 0.6 PDMS carry the conclusions: Table A.1 spans only 93.0 to 94.0 across 1 to 10 denoising steps, where 7 steps (94.0) beats the chosen 2 (93.8), and Table A.2 gives 93.2, 93.3 and 93.3 at 13, 15 and 17 modes. These defaults are also chosen on navtest itself. Significance testing is inexpensive here, since navtest is a fixed split and the evaluation emits per-scenario scores, so a paired test over scenarios or a bootstrap confidence interval requires no retraining at all.
  • Reproducibility. Hyperparameters like number of epochs, batch size and learning rate are not defined in Sec. 4.1.3.
  • Not clear how weight parameters (lambda, Sec. 4.1.3.) have been set. Ablation and tuning should be added.
Questions for authors
  • Which EPDMS does Table 2 use, the updated implementation that yields the higher scores, or the variant without human-behavior filtering that SparseDriveV2 reports separately? Please state this explicitly and evaluate every method under the same protocol.
  • Which inputs (agent futures, map, drivable area) does the PDM reranker in HTS use at inference?
  • How do you reconcile SCA’s +0.7 PDMS in Table 3 with its +3.3 in Table 6?
  • Please report paired per-scenario significance tests, or bootstrap confidence intervals, over the navtest split for every gain the paper claims, including the ablation differences of 0.1 to 0.6 PDMS on which the chosen defaults rest.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.