acceptodds
Withdrawn from ICLR 2027

SCRP: Repurposing Trajectory Prediction for Safety-Critical Relevance Planning with Agent Selection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Trajectory planning for autonomous driving requires reasoning about interactions among multiple agents under real-time constraints. Existing approaches often rely on increased model complexity and iterative trajectory generation to achieve strong performance, but incur high computational cost and limit scalability. In contrast, one-shot planners generate trajectories in a single forward pass and improve efficiency, yet fail to match iterative methods in complex scenarios. We identify that this performance gap arises from the absence of decision-aware interaction reasoning rather than limitations in model capacity or inference style. This paper presents a safety-critical relevance planning (SCRP) framework that enables efficient and high-performance one-shot planning. SCRP builds upon a generic state encoder–interaction encoder–trajectory decoder backbone inherited from trajectory prediction models and enables decision-aware interaction reasoning through two components: perceived safety modeling and risk-conditioned adaptive agent selection. A perceived safety modeling mechanism learns temporally evolving interaction risk and distills future-supervised safety cues into the latent interaction representation. In addition, a risk-conditioned adaptive selection mechanism allocates a scene-dependent interaction cardinality by retaining the most relevant agents under a risk-aware residual-mass criterion. The resulting interaction set adapts its capacity to scene complexity while suppressing weakly relevant context and preserving computational efficiency. Extensive experiments on large-scale autonomous driving benchmarks, including Waymo and nuPlan, demonstrate that SCRP consistently outperforms state-of-the-art (SOTA) planning models, particularly in challenging interactive and safety-critical scenarios. Adding SCRP reduces inference latency by 4.5%–9.1% relative to the corresponding prediction backbones, improving planning quality while retaining an efficient one-shot inference pipeline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous driving, Trajectory Planning, Repurposed Trajectory Prediction

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_JTcvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SCRP, a safety-critical relevance planning framework that adapts existing trajectory prediction models for autonomous driving planning. Its main components include a perceived safety prediction module, which learns future interaction risks through TTC, TET, and TIT indicators, and a risk-conditioned adaptive critical agent selection mechanism, which identifies relevant surrounding agents and dynamically determines how many should participate in ego-centric interaction modeling. The selected agents are then processed by the original interaction encoder and trajectory decoder to generate ego trajectories. For validation, the authors conduct experiments on WOMD and nuPlan using multiple prediction backbones. The results show that SCRP improves trajectory prediction accuracy and closed-loop planning performance while reducing inference latency.

Critical strengths
  1. SCRP provides a well-motivated perspective on repurposing prediction models for planning. Instead of increasing decoder complexity or relying on iterative generation, the method explores safety-critical relevance as a connection between trajectory prediction and ego planning.

  2. The experimental results are strong across a range of settings. SCRP improves several different backbones on both WOMD and nuPlan, with consistent gains in open-loop accuracy and closed-loop performance.

Critical weaknesses
  1. The use of predicted safety risk as a proxy for planning relevance is not fully validated. TTC/TET/TIT-based risk captures potential conflict, but an agent's influence on the ego decision can also arise from yielding, right-of-way, merging, or other interactions that do not necessarily correspond to high collision risk. The paper would benefit from further discussion of how these relevance scores reflect the agents that influence ego planning decisions.

  2. The interaction-risk modeling is not explicitly conditioned on candidate ego actions. SCRP predicts future TTC, TET, and TIT and uses these estimates to select relevant agents before generating the ego trajectory. However, the relevance of a surrounding agent may depend on the ego vehicle's intended maneuver. For example, an oncoming vehicle may be highly relevant when the ego intends to turn left, but irrelevant if the ego continues straight. It remains unclear how the proposed safety prediction and selection mechanisms account for such action-dependent interactions.

Questions for authors
  1. Could the authors further clarify how the proposed safety-aware relevance scores relate to actual planning relevance? In some interactions, such as yielding, merging, or right-of-way negotiation, an agent may influence the ego decision even when its immediate TTC/TET/TIT-based risk is relatively low. It would be helpful to understand how SCRP accounts for such cases when selecting relevant agents.

  2. Could the authors clarify how the ego vehicle's future action is reflected in the predicted interaction risk and the subsequent agent selection? Since the relevance of a surrounding agent may depend on what the ego vehicle plans to do, it would be helpful to understand whether and how the safety branch accounts for this dependence.

Reviewer_Uc7aOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the problem of generating safe and feasible ego trajectories from historical agent states and map context under real-time constraints in multi-agent driving scenes. It proposes SCRP, a framework that adapts trajectory prediction backbones to planning by identifying safety-relevant interactions while retaining one-shot inference. A perceived-safety branch predicts future TTC/TET/TIT sequences and produces safety-aware features. A risk-conditioned selector combines feature similarity and predicted risk to retain a scene-dependent number of agents under a residual relevance-mass criterion. The selected features are processed by the backbone’s interaction encoder and trajectory decoder. Experiments on WOMD and nuPlan report improvements across multiple backbones and modest inference-latency reductions. The evaluation includes safety-component ablations, fixed-versus-adaptive selection comparisons, and a conditional bounded-omission analysis.

Critical strengths
  1. Consistent improvements across different backbones.

The reported closed-loop scores improve across all four evaluated backbones and all four nuPlan evaluation settings. This breadth suggests that the proposed additions may offer benefits beyond a single architecture.

  1. A practical approach to improving planning without replacing the backbone.

SCRP retains the core interaction encoder and trajectory decoder while introducing safety supervision and selective interaction processing. The reported 4.5%–9.1% latency reductions relative to the corresponding prediction backbones suggest a potentially useful combination of improved planning quality and reduced inference cost.

Critical weaknesses
  1. Limited empirical support for the central diagnosis and attribution of improvements.

The Introduction attributes the performance gap primarily to missing decision-aware interaction reasoning rather than model capacity or one-shot inference. However, no diagnostic or controlled experiment establishes this explanation. The reported gains demonstrate the usefulness of the proposed additions, but do not exclude alternative explanations such as auxiliary supervision or generic interaction filtering. The existing ablations also leave important mechanisms entangled: Table 3 removes the entire safety branch, affecting both representation conditioning and risk-guided selection; Figure 3 examines safety-input channels; and Appendix H compares fixed and adaptive cardinalities. These experiments do not isolate the two safety-information pathways or establish whether future-safety supervision is more beneficial than a comparable auxiliary prediction task. Consequently, it remains difficult to determine which proposed mechanism accounts for the gains.

  1. Unclear novelty and weak justification of the relevance mechanism.

Prediction-inspired planning and integrated prediction–planning are already well explored. BeTop, one of the adopted backbones, already incorporates future-interaction supervision and selective interaction modeling. The paper therefore needs a clearer account of its technical advance beyond these approaches; repurposing prediction architectures alone does not establish a new perspective. Moreover, the core selector assumes that a smaller Euclidean distance between fused ego and agent embeddings indicates greater planning relevance. Feature similarity does not inherently imply influence on the ego decision, and the experiments do not directly validate this relationship or compare alternative relevance functions. The omission bound is conditional on the estimated relevance approximating an unknown influence distribution; it does not establish that the proposed distance-based score satisfies this condition. These gaps limit my assessment of both the contribution’s originality and the necessity of its specific design.

Questions for authors
  1. What empirical evidence supports the claim that missing decision-aware interaction reasoning, rather than model capacity or inference style, is the principal source of the performance gap? How is this reasoning capability operationally defined and measured?

  2. Can you provide controlled comparisons using safety information only for representation conditioning, only for agent selection, for both, and for neither? A comparison against a similarly sized auxiliary future-prediction task would also help determine whether the gains are specific to safety supervision.

  3. Why should Euclidean distance between ego and agent embeddings reflect decision relevance? How are the features aligned, fused, and scaled in Eq. (3)?

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

This paper proposes SCRP, a framework that repurposes trajectory prediction backbones for efficient one-shot planning through safety-aware interaction modeling and adaptive agent selection. It introduces future-supervised safety representations and risk-conditioned agent selection to identify planning-relevant agents, demonstrating substantial improvements across multiple backbones on WOMD and nuPlan in both open-loop and closed-loop evaluations.

Critical strengths

S1. The paper presents a valuable insight into improving planning through safety-critical interaction relevance rather than more complex trajectory generation. This offers a meaningful perspective on bridging the gap between trajectory prediction and planning while preserving the efficiency of one-shot inference.

S2. The proposed framework demonstrates strong and consistent empirical improvements across multiple prediction backbones. The substantial gains in planning performance, together with reduced inference latency, provide convincing evidence of its practical effectiveness and architectural generality.

Critical weaknesses

W1. The evaluation insufficiently establishes the distinctiveness and effectiveness of the proposed agent selection mechanism. Although fixed-cardinality comparisons are provided, stronger baselines using alternative relevance-based selection strategies are missing, and there is limited direct evidence that the selected agents are indeed the most influential for ego planning.

W2. The contributions of safety modeling, adaptive selection, and training adaptation are not sufficiently disentangled. While the paper includes component ablations, the independent contributions and interactions of the proposed mechanisms remain unclear, and the training protocols do not clearly establish how much of the improvement is attributable to SCRP rather than backbone fine-tuning or task adaptation.

Questions for authors

Q1. How does RCAS compare with alternative agent selection strategies under matched interaction budgets and training settings? Comparisons with distance-based, attention-based, or risk-only selection, together with analyses of the planning impact of removing selected agents, would help establish whether RCAS identifies genuinely decision-relevant agents.

Q2. Can the authors clarify the backbone training protocol and further disentangle the sources of performance improvement? In particular, are backbone parameters fine-tuned, and how do safety-only, selection-only, and full SCRP variants perform under identical training conditions?

Q3. Is the benefit of SCRP specific to one-shot planning, or could similar improvements be expected for iterative planners? Clarifying this would help assess whether SCRP addresses a limitation specific to one-shot planning or introduces a more general interaction selection mechanism.

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

Related papers

Discussion (0)

Sign in to comment.