acceptodds
Under review as a conference paper at ICLR 2027

Safe Scenario Parameter Region Synthesis for End-to-End Autonomous Driving

Abstract

End-to-end autonomous driving models have demonstrated great potential in driving safely under normal conditions. Existing techniques that evaluate their behavior on long-tail scenarios focus on generating diverse scenarios to test the driving system. In this work, we characterize safety-critical scenarios by certain parameters, and study the synthesis of the set of values for these parameters under which the Autonomous Vehicle is safe. We formalize the problem of synthesizing safe and maximal scenario parameter regions, and propose a framework to solve it. Our framework, Odin, uses a counterexample-driven algorithm to achieve its objective. We evaluate Odin on safety-critical scenarios for its efficacy in synthesizing safe regions and producing safe/unsafe pairs of parameters using the regions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, Safe region synthesis

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_F7mPOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how to identify safe regions of scenario parameters for end-to-end autonomous driving models. The authors propose ODIN, a framework that uses decision trees to learn statistically safe and locally maximal parameter regions. ODIN runs driving simulations to check the learned regions and iteratively updates the decision trees using safe and unsafe samples. The authors evaluate ODIN on seven safety-critical scenarios in CARLA using the TransFuser model. The results show that ODIN converges across different decision tree depths and can find pairs of safe and unsafe scenarios near the learned region boundaries.

Critical strengths
  1. The paper studies a meaningful problem in autonomous driving safety evaluation. Learning safe parameter regions can help understand the conditions under which an autonomous driving model behaves safely. The boundaries of these regions can also provide useful safe and unsafe scenario pairs for analyzing model failures.

  2. The proposed framework provides an interpretable way to describe safe parameter regions using decision trees. The authors also evaluate the framework on seven different driving scenarios in CARLA. The results show that ODIN can converge across different tree depths and identify safe and unsafe configurations near the learned boundaries, which supports the feasibility of the approach.

Critical weaknesses
  1. The first concern is that the advantages of the proposed framework over other approaches are unclear. Each scenario is represented by a small number of parameters, and the simulator provides safe or unsafe labels for sampled configurations. It seems possible to directly sample the parameter space and learn the safe/unsafe boundary using a classifier or a prediction model. However, the paper does not compare ODIN with such approaches. It also does not evaluate the coverage or boundary accuracy of the learned safe regions. Without these comparisons, it is difficult to understand the necessity and practical benefits of the proposed framework.

  2. The second concern is the limited experimental evaluation. ODIN learns a separate safe region for each predefined scenario, and the experiments only consider one autonomous driving model. It is unclear how useful these learned regions would be when the driving scenarios or models change. Therefore, the current experiments provide limited evidence of the broader usefulness of the proposed approach.

Questions for authors

I think the authors should provide more analysis and evidence to demonstrate the value of the proposed framework. Specifically:

  1. Could the authors provide additional comparisons or analyses to show its benefits in terms of safety region coverage, boundary accuracy, or simulation efficiency?

  2. Could the authors clarify whether ODIN needs to learn a new safe region from scratch when the driving scenario or autonomous driving model changes? How practical is this approach when applied to a large number of scenarios or different driving models?

2 comments
AuthorsOpenReview ↗

Response to Reviewer F7mP

We thank the reviewer for their comments and address their main concerns below:

Could the authors provide additional comparisons or analyses to show its benefits in terms of safety region coverage, boundary accuracy, or simulation efficiency?

Yes, we will provide in the final version, a baseline that samples configurations randomly from the entire space, and learns trees for a bounded number of samples/iterations, followed by testing it for safety and maximality, similar to the current evaluation. This will help understand the effect of guiding the sampling using counterexamples and focusing on safety and maximality testing.

Could the authors clarify whether ODIN needs to learn a new safe region from scratch when the driving scenario or autonomous driving model changes? How practical is this approach when applied to a large number of scenarios or different driving models?

Yes, ODIN learns a new safe region for each driving scenario and model. Learning a safe region from scratch is necessary for different models, of course, because they can exhibit different safety regions. The intent behind ODIN is that practitioners can use it on representative scenarios that are relevant to their deployments and get an understanding of their model's behavior in different configurations of a scenario. However, when models evolve (e.g., improve), we see a path to extend ODIN to incrementally build safe regions, evaluating safety and maximality (but with the same set of anchor points), and continuing the learning algorithm. We leave this for future work.

Thank you for the clarifications. However, my main concerns remain unresolved.

Reviewer_JkcYOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors propose an approach to synthesize regions having the safe autonomous driving behavior. They did it using the specialized framework for deciding which parameter regions are safe and which are not.

Critical strengths
  • Usage of decision trees over inequalities to construct the rules of safety
  • Maximal region that are still safe - as an approach to construct of the trees (see above)
Critical weaknesses

The main question is the following: how to actually use the knowledge of what is safe and what is not? How to enrich the training datasets in a way that targeting the hardest long-tail cases now makes much more sense? How to really evaluate the impact of the proposed solution - maybe all rules the are just hypothetical constructions? There is nothing covering this part in the paper, and I do think this is super important for such a paper to be able to prove the importance of the safety/non-safety regions.

Testing only using CARLA and the Latent TransFuser v6 is just simply not enough for any statistical conclusion.

Questions for authors

I'm not sure how to objectively measure the impact of the constructions.

2 comments
AuthorsOpenReview ↗

Response to Reviewer JkcY

We thank the reviewer for their comments and address their main concerns below:

How to really evaluate the impact of the proposed solution - maybe all rules are just hypothetical constructions? There is nothing covering this part in the paper, and I do think this is super important for such a paper to be able to prove the importance of the safety/non-safety regions.

We don’t understand this point. Our regions are provably maximal in the sense that their expansions have unsafe points. And we sample configurations from the safe regions to test the safety of the region, and show that most of the learned trees accurately identify safe configurations. Safety here is defined as a function of the simulator's state, and is checked independent of the ODIN algorithm.

If you mean the utility of the safe regions in end-to-end driving models, the focus of this paper is on defining and studying the novel problem of finding safe scenario parameter regions. While we argue that these learned regions have potential in serving downstream purposes (e.g., retraining the driving model on unsafe scenario configurations, comparing the quality of two driving models using their safe regions, etc.), the paper does not establish utility for these purposes. We consider this out of scope for this paper, and part of future work which evaluates the application of safe regions in larger commercial settings.

Testing only using CARLA and the Latent TransFuser v6 is just simply not enough for any statistical conclusion.

We conduct statistical tests of the learned safe regions and report the outcomes in the paper. We do not consider all combinations of driving models and driving simulators to perform statistical tests. We will add a threat to validity where we acknowledge that our results are not shown to generalize across models.

Thanks for your comment. Anyway, it doesn't resolve my concerns so I'm keeping my score.

Reviewer_zjk3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ODIN, a counterexample-guided framework for learning interpretable safe regions in scenario parameter space for end-to-end autonomous driving. A decision-tree learner alternates with simulation-based safety and expansion tests, with a discretization-based conflict resolver modifying training labels. Experiments use Latent TransFuser v6 in CARLA on seven scenarios and report convergence, post-hoc statistical checks, and illustrative safe/unsafe configuration pairs.

Critical strengths
  1. Learning interpretable safe regions complements failure discovery and can help explain parameter changes that alter driving outcomes.
  2. The learner-tester architecture treats the driving model as a black box and requires neither gradients nor access to internal representations.
  3. The seven scenarios cover interactions with pedestrians, cyclists, and vehicles; the case studies illustrate understandable changes in weather and actor behavior.
Critical weaknesses
  1. The practical value of synthesizing safe regions is not established. The paper motivates these regions as a source of boundary scenarios for improving driving models, but provides no retraining experiment. It is unclear whether this formulation yields more useful training data than ordinary failure discovery. At least one controlled fine-tuning study should measure held-out closed-loop safety and driving performance under equal data and training budgets. Without such evidence, the downstream value remains speculative and the motivation for this particular region-synthesis objective is insufficient.
  2. Statistical safety is checked after termination rather than enforced during synthesis. Definition 4 requires a confidence lower bound at least as large as a specified threshold, but Algorithm 1 has no corresponding acceptance criterion. Table 1 reports 90% confidence lower bounds of 0.59 for the depth-5 curbside scenario and 0.65 for depth-6 cyclists, below the stated 0.7 target. Although another depth passes in each scenario, convergence itself does not enforce the definition. Integrate independent certification and specify how failed checks are handled. The p0 = 0.7 choice also needs sensitivity analysis below and above 0.7, including stricter targets, using sufficient independent samples. Report changes in certifiability, coverage, simulation cost, and downstream training gains.
  3. The implementation does not establish anchoring and conformance. As written, Algorithm 3 relabels any safe point sharing a discretization bucket with an unsafe point, without protecting initial anchors. This can contradict Definition 4, which requires retaining every initial positive point. The depth-limited scikit-learn learner is also not shown to exclude all known negatives. Explain how these constraints are enforced, including cases where the depth limit makes them infeasible. Clarify how exact conformance on observed samples, population-level statistical safety, and conservative relabeling fit together.
  4. Maximality testing and convergence need a consistent account. Section 4 accepts an expansion containing at least one unsafe witness, whereas Algorithms 1-2 require agreement on every sampled label; a safe sample outside the predicate triggers refinement even when an unsafe witness exists. The pseudocode also omits the stated exclusion of previously ignored counterexamples. Reconcile these descriptions and provide correctness and termination conditions. Maximality relative to the chosen Delta expansion should be distinguished from identifying the entire safety frontier. Ablate the policy of stopping an expansion after one unsafe witness against risk-based acceptance or finer local refinement, and measure effects on boundary quality and subsequent driving-model training.
  5. No controlled comparisons establish the benefit of the proposed components. Table 1 varies tree depth but includes no baseline with a matched simulation budget. Compare against passive uniform sampling with the same tree learner and simple boundary-focused active sampling. Ablate maximality queries, conflict resolution, and the epsilon gap individually. Convergence and selected qualitative pairs alone do not establish the value or efficiency of these components. Evaluate alternative boundary-finding methods by the gains they enable after the same driving-model fine-tuning protocol, with matched simulation, selected-data, and training budgets and held-out closed-loop evaluation.
  6. Region quality and robustness are insufficiently measured. High safety precision can result from predicting a small region, yet region volume, safe-point recall, and boundary-pair yield or distance are not quantified. Evaluation uses one driving model and at most six active parameters, with others restricted to canonical safe ranges. Repeated runs, larger independent test sets, and explicit simulator-seed handling would clarify variability and the scope of the safety claim.
Questions for authors
  1. How are failed statistical checks handled, and how do region coverage, simulation cost, and training gains change with safety thresholds below and above 0.7?
  2. Can Algorithm 3 relabel initial anchors, and how are all anchoring and negative-exclusion constraints enforced with a depth-limited tree?
  3. How do results change when one unsafe witness no longer stops an expansion? Please ablate this rule and reconcile the implemented maximality criterion with its definition.
  4. Which boundary-finding method yields the largest held-out improvement after driving-model fine-tuning under matched simulation, selected-data, and training budgets?
2 comments
AuthorsOpenReview ↗

Response to Reviewer zjk3

We thank the reviewer for their comments and address their main concerns below:

The practical value of synthesizing safe regions is not established. The paper motivates these regions as a source of boundary scenarios for improving driving models, but provides no retraining experiment.

The focus of this paper is on defining and studying the novel problem of finding safe scenario parameter regions. While we argue that these learned regions have the potential to serve downstream purposes (e.g., retraining the driving model on unsafe scenario configurations, comparing the quality of two driving models using their safe regions, etc.), the paper does not establish utility for these purposes. We consider this out of scope for this paper, and part of future work which evaluates the application of safe regions in larger commercial settings.

Statistical safety is checked after termination rather than enforced during synthesis.

The statistical check is used during the synthesis, but in an outer loop. In our experiments, if the statistical safety check fails, the learning loop is repeated up to a fixed number of repetitions. If a statistically safe region is not learned in those repetitions, we declare the learner to have failed to learn a safe region. Such scenarios occur when the safe region for a scenario is inexpressible in the given tree depth. We will clarify the presentation by adding the statistical safety check in the algorithm presented in the Appendix.

The implementation does not establish anchoring and conformance.

The algorithm is indeed required to be faithful with respect to the anchor points. It will fail to learn a safe region if an anchor point is flipped to unsafe.

Maximality testing and convergence need a consistent account. Section 4 accepts an expansion containing at least one unsafe witness, whereas Algorithms 1-2 require agreement on every sampled label; a safe sample outside the predicate triggers refinement even when an unsafe witness exists.

Each candidate counterexample for maximality represents one possible expansion of the safe region, within the specified δ\delta-expansion limits. If one such candidate turns out to be a counterexample, the algorithm explores the expansion of the safe region that contains the maximality counterexample.

Ablate maximality queries, conflict resolution, and the epsilon gap individually. Repeated runs, larger independent test sets, and explicit simulator-seed handling would clarify variability and the scope of the safety claim.

Performing all the ablations would take at least 2619 uninterrupted hours (over 109 days) on our clusters (estimated from the current experiment durations; one ablation for maximality, one for conflict resolution, two for different epsilon gaps, two repeated runs, two simulator seeds, and a 2x larger independent test set to evaluate the trees). These ablations would be too expensive to perform (in terms of time and computational resources).

How are failed statistical checks handled?

If a safe region fails a statistical check, ODIN tries to re-learn a safe region in a loop with a fixed loop budget.

Can Algorithm 3 relabel initial anchors, and how are all anchoring and negative-exclusion constraints enforced with a depth-limited tree?

Addressed above.

How do results change when one unsafe witness no longer stops an expansion? Please ablate this rule and reconcile the implemented maximality criterion with its definition.

Addressed above.

Which boundary-finding method yields the largest held-out improvement after driving-model fine-tuning under matched simulation, selected-data, and training budgets?

We do not perform fine-tuning in this work. Our proposed definition of safe scenario parameter regions can be used by practitioners in the autonomous driving industry to fine-tune their models (which are not available to us) or use the regions to understand the failure modes of their models.

Thank you for your response. However, I find that some of my concerns have not been sufficiently addressed. Therefore, I will maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.