acceptodds
Under review as a conference paper at ICLR 2027

TraffusionGen: Parallel Fusion Framework for WaymoTraffic Simulation

Abstract

High-fidelity traffic scenario generation is a cornerstone for safe and scalable autonomous driving validation, enabling closed-loop policy learning, rare-case testing, and low-cost data augmentation. Existing data-driven methods suffer from three critical limitations: (1) fragile trajectory completion under occlusion and sparse observations, (2) inconsistent long-term multi-agent coordination, and (3)limited behavioral diversity prone to mode collapse. To address these challenges,we propose TraffusionGen, a unified parallel fusion framework that explicitly decouples and integrates global contextual constraints, multi-agent interactive dependencies, and stabilized residual feature fusion within a single end-to-end architecture.Our model introduces three core innovations: (1) a unified vectorized scene encoder that jointly embeds dynamic agents, lane topology, and semantic map signals into a shared representation space; (2) a novel Parallel MCG-Attention Block that processes global context guidance and local multi-agent self-attention in parallel, eliminating representational bias in conventional sequential designs; (3) a context-conditioned autoregressive decoder that unifies scenario generation, missing trajectory completion, and interactive scenario augmentation under consistent physical and semantic constraints.Extensive experiments on the Waymo Open Motion Dataset demonstrate that TraffusionGen achieves state-of-the-art performance across four challenging settings:standard generation, vehicle-masked completion, traffic-light-masked completion,and map-vector restricted completion. Our method significantly outperforms baselines including TrafficGen, UniGen, LCTGen, and SceneGen in distribution fidelity (MMD), trajectory accuracy (ADE/FDE), and physical safety (collision rate).TraffusionGen establishes a new generic backbone for high-fidelity traffic simulation and provides principled insights for future research in data-driven autonomous driving simulation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
traffic scene generation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_tVvJOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a traffic scene simulation method named TraffusionGen. This method treats generation and completion as task-conditioned scene inference. In order to achieve better results, it combines the MCG and self attention operation independently on the same incoming tokens, using the learnable residual connections to model the scene-level context and token-level relationships. On the processed WOMD subset, this method improves the generation metrics and maintains lower ADE under vehicle masking. Component ablations support gains from both branches and their fusion.

Critical strengths
  1. Parallel MCG-Attention Blocks combine the MCG and self attention operations to get better results compared with other baseline methods. As shown in Table2, TraffusionGen obtains 0.31 m ADE and 0.40 m FDE, reducing TrafficGen baseline values by 46.6% and 46.7%. Position MMD decreases from 0.35 to 0.12, and collision rate from 0.38 to 0.18. These consistent improvements across multiple evaluation metrics demonstrate that combining MCG with self-attention enables TraffusionGen to achieve more accurate trajectory generation, better distribution matching, and fewer collisions in generated traffic scenes.
Critical weaknesses
  1. It would have been better to add explainations of the motivation to combine the MCG and self attention operation. These operations are common, and used by many previous works. This paper seems to add these two operations together simply.
  2. The paper don't provide the details of network architectures and training hyper parameters. For example, the architecture details of PolylineEncoder, EmbedSemantic, and the decoder. It's hard for other researchers to reproduce this results. Providing these details would make this work reproducible.
Questions for authors
  1. Provide more details of network architectures and training hyper parameters to help reproduce this work. Providing these details would enable other researchers to reproduce and verify the reported results.
  2. Motivation clarification of MCG and self attention combination. Why their combination is particularly suitable for traffic scene generation and completion, and what specific limitation of existing architectures this parallel design is intended to address.
  3. Add a comparision on other dataset to validate the robustness of proposal method. TraffusionGen was only tested on WOMD subset.
  4. Add some visualization comparison results to help reader understand. This paper don't provide any visualization results.
Reviewer_zzEcOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents TraffusionGen, a framework that treats generation and completion as task-conditioned scene inference. Agents, lanes, and semantic map elements are encoded as shared scene tokens. A Parallel MCG-Attention Block reads the same tokens through two branches: multi-context gating broadcasts shared scene context, while self-attention computes query-specific relational updates. Learnable residual fusion combines these updates before a context-conditioned autoregressive decoder predicts the requested variables.

Critical strengths

None

Critical weaknesses
  1. Insufficient Novelty and Motivation
    The parallel MCG-Attention relies on existing components (MCG from MultiPath++ and standard self-attention), and parallel plus residual fusion is also fairly common. The authors should further explain why this design outperforms cross-attention or explicit gating, and add corresponding ablation studies.

  2. Dataset Choice and Generalization
    The experiments use only a filtered subset of WOMD, which may miss long-tail interactions and lead to overfitting to a specific distribution. Evaluation on the full WOMD, or an analysis of how filtering affects generalization, is recommended.

  3. Lack of Sufficient Evaluation
    The open-loop metrics are solid, but a paper centered on scene generation also needs diversity metrics and closed-loop evaluation in simulators (e.g. CARLA). Adding these would improve practical relevance.

Questions for authors

Please refer to the Weaknesses.

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

This paper proposes TraffusionGen, a unified framework for traffic scene generation and completion under partial observations. The method introduces a parallel architecture that combines Multi-Context Gating (MCG) for global scene context modeling with multi-head self-attention for token-level interactions, followed by learnable residual fusion and autoregressive decoding. Experiments on a processed subset of the Waymo Open Motion Dataset demonstrate improvements over SceneGen, TrafficGen, and UniGen, achieving a 46.6% reduction in ADE compared with the strongest baseline and showing greater robustness to missing vehicle observations. Ablation studies further indicate that both context modeling branches and their fusion contribute to the reported performance improvements.

Critical strengths
  1. The proposed architecture integrates traffic scene generation and completion within a shared framework.

  2. The method demonstrates substantial improvements over the selected baselines across multiple evaluation metrics.

Critical weaknesses
  1. Limited Methodological Novelty. The proposed method primarily combines existing multi-context gating (MCG), standard multi-head self-attention, and learnable residual fusion within an autoregressive traffic scene generation framework. Since MCG-based context modeling and attention-based traffic generation have already been explored in prior work, the methodological contribution appears largely incremental. Moreover, the paper provides limited justification for why the proposed parallel fusion is preferable to simpler sequential or alternative fusion architectures, and the current ablations do not sufficiently isolate the benefits of this specific design.

  2. Limited Comparisons with Recent Methods. The experimental comparisons are limited to SceneGen (2021), TrafficGen (2022), and UniGen (2024), overlooking more recent traffic scene generation methods such as GUMP (ECCV 2024) and UMGen (CVPR 2025). Although the paper discusses other recent approaches, it does not adequately justify their exclusion from the evaluation.

  3. Limited Experimental Scope and Generalization Evidence. All experiments are conducted on a single processed subset of the Waymo Open Motion Dataset, with only 1,000 test scenarios. The absence of cross-dataset evaluations, more diverse traffic conditions, and repeated runs limits the evidence for generalization and robustness. Given the substantial improvements reported over existing baselines, more comprehensive experiments and variability analysis would be necessary to substantiate the claims.

  4. The paper does not clearly distinguish between reconstructing entirely missing agents and completing partially observed trajectories. Moreover, Figure 3 reports a nonzero ADE at 0% vehicle masking, although the completion task is described as predicting only masked targets. The authors should clarify the exact prediction targets, masking procedure, and evaluation protocol across different masking ratios.

  5. Although the paper presents a unified framework for traffic scene generation and completion, it does not demonstrate whether joint training provides advantages over independently trained task-specific models. The absence of such comparisons makes it difficult to assess the practical benefits of the proposed task unification beyond architectural reuse.

Questions for authors

None

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.