acceptodds
Under review as a conference paper at ICLR 2027

CuraScene: Interaction-Aware Scenario Curation for Data-Efficient Autonomous Driving

Abstract

Trajectory forecasting models for autonomous driving are trained on scenario corpora that now reach hundreds of thousands of samples, yet the marginal value of an additional scenario is far from uniform. We argue that much of this redundancy is structural: what makes a driving scenario informative is largely the pattern of predictive influence among its agents—whose past states help forecast whose futures—and large corpora repeat a small number of interaction structures while varying mostly in surface trajectory geometry. We present CuraScene, which makes this structure explicit and selects on it. For each scenario, CuraScene infers—without interaction supervision—a directed graph over the agents and, where annotated, the traffic signals that control them. An edge indicates that the lagged state of improves prediction of . The graph is learned with an encoder–decoder trained under a sparsity penalty and a physically admissible edge mask. Scenarios are then selected greedily by the topological novelty of their ego-centric subgraph relative to the already-selected set, so a fixed budget covers interaction structures rather than raw samples. We do not claim a new graph structure learning algorithm; our claim is that inferred interaction structure is a useful and inexpensive selection signal for driving data. At a matched 50% budget on Waymo Open Motion and Argoverse 2, two forecasting models trained on CuraScene-selected scenarios reduce by 10.7–14.6% relative to budget-matched random subsampling, and an ablation shows that the learned graph and its admissibility prior are both required in order to recover the agents Waymo annotates as interacting.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Data Curation, Data Selection, Autonomous Driving, Multi-Agent Learning, Graph Learning, Representation Learning, Machine Learning Systems

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_86DUOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

CuraScene selects training scenarios for motion forecasting using inferred interaction graphs. It learns sparse directed graphs from agent trajectories, extracts ego-centric subgraphs, and selects scenarios according to their structural novelty. At a reported 50% data budget, the selected subsets improve SeNeVa and DeMo over random subsampling on Waymo and Argoverse 2. The paper also evaluates agreement between the recovered agent sets and a relevance proxy derived from Waymo annotations.

Critical strengths

The problem is practically relevant. Using interaction structure to identify redundant driving scenarios is an interesting idea, and the selection method can be used without changing the downstream forecasting model.

Critical weaknesses

The experiments do not establish the main claimed contribution. Random sampling is the only selection baseline, and all downstream results use a single 50% budget. There is no comparison with embedding-based diversity selection or a simple interaction-density heuristic. More importantly, Table 2 is not a downstream selection ablation: it measures agreement with a set of relevant agents, rather than forecasting accuracy after training on subsets selected by the ablated methods. It also does not evaluate edge direction. Consequently, the results do not separate the value of learned directed topology from ordinary deduplication or preferential selection of more interactive scenes. Since the paper explicitly places its contribution in the selection signal rather than a new graph-learning algorithm, this missing evidence weighs heavily in my assessment.

The matched-budget experiment cannot be fully reconstructed from the reported protocol. Section 5.1 states that both methods select exactly half of the original scenarios. Algorithm 1, however, first removes duplicate graphs and then calibrates approximately 50% retention, without explaining how this produces the required number of original scenarios. The Argoverse 2 split also needs clarification: Appendix B describes the 224,896 scenarios as the combined training and validation splits, while the main experiment uses this number to describe the pool without identifying the held-out evaluation set. Actual split and subset counts are needed to verify the comparison. The paper acknowledges several limitations, but these details concern the validity of its main result.

The comparison with existing work is not convincing enough to establish the contribution. Most of the related-work discussion focuses on work published in 2024 or earlier, and the evaluation uses SeNeVa and DeMo, both from 2024. More importantly, these are downstream forecasting models; the only baseline for the proposed selection strategy is uniform random sampling. Beating random sampling does not establish an advantage over existing diversity-based or coreset selection methods. Given that the graph-learning and selection components largely build on established techniques, this missing comparison makes it difficult to assess the novelty and practical value of their combination. The authors should engage more closely with recent relevant work, including applicable 2026 methods available before submission, and justify their choice of baselines. As presented, the evidence for an advance over competitive selection approaches is insufficient.

I lean toward rejection. The gains over random sampling are promising, but they do not yet establish the value of the proposed interaction-based selection criterion. The missing downstream controls and the unclear budget and split accounting concern the paper’s main evidence, so I would need these issues resolved before supporting acceptance.

Questions for authors

Please provide the Argoverse 2 split sizes used for graph training, scenario selection, downstream training, and evaluation. After graph deduplication, how many scenarios remain, and how is the exact 50% budget enforced for both methods?

What node attributes and matching constraints are used in canonicalization and graph edit distance? In Appendix E(b), exchanging two otherwise indistinguishable agents A and B leaves a rooted graph with one incoming and one outgoing edge isomorphic to the original. Under a topology-only distance, how are the two examples distinguished?

Are any downstream results available for selection using the heuristic prior alone, undirected graphs, or generic feature diversity? Please also report the number of repeated runs and the variability behind Table 1.

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

This paper proposes a dataset selection method for autonomous driving motion prediction that infers directed agent interaction graphs in an unsupervised manner and performs greedy subset selection based on interaction topology novelty. Under the same data budget, the method prioritizes scenes with diverse interaction structures, thereby improving the training performance of downstream motion prediction models compared with random selection.

Critical strengths

The paper introduces a novel data valuation approach. Unlike conventional methods relying on explicit scene features, it employs a custom encoder–decoder architecture to capture implicit inter-vehicle influences.

Critical weaknesses
  • Insufficient experimental evaluation. Acknowledging limitations does not compensate for missing experiments. The paper lacks comparisons with diverse data selection baselines and evaluations across different selection ratios, making its advantages difficult to assess. Moreover, the encoder–decoder's prediction accuracy and qualitative examples of inferred interaction graphs are not provided.
  • Unjustified architectural decoupling. The complete separation between the encoder–decoder and downstream prediction models lacks justification. While this design offers generalizability, potential integration into downstream models remains unexplored.
  • Poor readability. The manuscript contains excessive long sentences, qualifications, and defensive statements. Key concepts, such as normalized GED, ego-centric subgraphs, and nonlinear Granger predictive influence, lack intuitive examples, making the paper unnecessarily difficult to follow.
Questions for authors
  • Impact of binary graph representation. The selection method uses binary unweighted graphs, discarding edge confidence and interaction strength. Could this substantially affect selection quality, given the significant differences between strong and weak interactions?
  • Insufficient experimental evidence. Acknowledging limitations cannot replace experimental validation. Could the authors provide comparisons with additional data selection baselines, results across multiple selection ratios, prediction accuracy metrics for the encoder–decoder, and examples of inferred interaction graphs?
Reviewer_5KddOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CuraScene, a data selection method that prioritizes diversity in inferred interaction structures. It learns directed graphs from agent trajectories without interaction labels and selects scenarios based on the novelty of their ego-centric subgraphs. Experiments with two forecasting models on Waymo Open Motion and Argoverse 2 report improved prediction accuracy over random selection at a matched data budget.

Critical strengths
  1. The paper addresses a useful practical problem by using inferred interaction structure to select training scenarios. Consistent forecasting improvements over budget-matched random selection across two datasets and two models provide initial evidence that this representation can support effective data curation.
Critical weaknesses
  1. The evaluation compares only against random subsampling, leaving the main contribution insufficiently validated. Without comparisons to diversity-based selection in trajectory or scene embedding spaces, it is unclear whether inferring interaction graphs provides additional value over established selection criteria. This comparison is particularly important because the paper’s contribution centers on the selection representation.
  2. The contribution of graph learning and edge directionality to the reported forecasting gains is unclear. Table 2 evaluates how well graph variants identify relevant agents, but does not test the forecasting accuracy of models trained on the subsets selected by each variant. In particular, matched-budget forecasting comparisons using the heuristic graph or an undirected version of the learned graph are missing. These comparisons are needed to establish whether learning directed interactions provides additional value for data selection.
Questions for authors
  1. How does CuraScene compare with representative existing data curation methods under the same data budget and forecasting-model training settings? Can these comparisons establish the added value of using interaction structure as the selection criterion?
  2. Could you report forecasting results from models trained under identical conditions on equally sized subsets selected using the heuristic graph in Table 2 and an undirected version of the learned graph? These comparisons would help clarify the respective contributions of graph learning and edge directionality to the forecasting gains.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.