acceptodds
Under review as a conference paper at ICLR 2027

Colligen: Physics-Guided Self-Learning for Generating Collision Trajectories of Vehicles

Abstract

Collision data are critical for improving the safety and robustness of autonomous driving systems. However, real collision trajectories are scarce and expensive to collect. We address this limitation by generating physically plausible collision tra- jectories using pretrained normal-driving predictors from abundant normal-driving data. We propose COLLIGEN, a physics-guided self-training framework that pro- gressively adapts a normal-driving trajectory model into a collision trajectory generator. COLLIGEN consists of two training phases. In the Pre-collision phase, the model learns trajectories that lead to impact while remaining behaviorally plausible; in the Post-collision phase, it learns trajectories that satisfy post-impact physical constraints. To guide the adaptation, we adopt Markov Chain Monte Carlo sampling with physics-based scoring to enhance collision feasibility, dynamic consistency, and motion regularity of self-generation. Experiments show that COL- LIGEN substantially outperforms strong text-to-trajectory, trajectory-to-trajectory, and trajectory-prediction baselines in reconstructing collision trajectories synthe- sized from NHTSA accident reports. Qualitative results show that COLLIGEN produces more realistic and physically plausible pre- and post-collision motion.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, Data Generation, Physical AI

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xZMwOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

ColliGen adapts normal-driving trajectory predictors into collision trajectory generators through two-phase, physics-guided self-training. Each phase alternates between generating pseudo-trajectories and fine-tuning the predictor. Bézier-based perturbations and independent Metropolis–Hastings sampling guide candidate generation, while physical scores encourage collision feasibility and plausible post-impact motion. The method is evaluated using HiVT and QCNet on 4,583 accident scenarios reconstructed from NHTSA records with CARLA/Chrono. Reported displacement errors improve substantially over the compared baselines, particularly after impact.

Critical strengths
  1. Meaningful problem and well-motivated adaptation framework. Complete collision trajectories are scarce, and adapting predictors trained on abundant normal-driving data is a useful direction. The Grow–Improve procedure and the separation of pre- and post-collision training provide a coherent approach to progressively learning these rare events.

  2. Substantial physical scoring design with encouraging empirical results. The framework combines collision feasibility, post-impact momentum and velocity consistency, and motion regularity. The reported improvements are considerable, and the scoring ablations support the usefulness of combining impact physics with motion regularity.

Critical weaknesses
  1. The central physical-plausibility claim lacks direct quantitative validation. Evaluation primarily measures displacement and relative-distance errors, which do not establish collision success, absence of interpenetration, or physically consistent impact responses.

  2. The practical utility of the generated data remains unclear. Autonomous-driving training and evaluation motivate the work, but no downstream experiment establishes whether the generated trajectories improve a predictor or driving policy, or provide effective safety tests. Reconstruction accuracy alone does not substantiate these intended uses.

Questions for authors
  1. Could the authors directly evaluate the physical validity of the generated trajectories, including collision success rates, interpenetration, and post-impact momentum or impulse consistency?

  2. Could the authors demonstrate the utility of the generated data through a focused downstream training or safety-testing experiment, analogous to the policy-training evaluation in TrafficGen [1]? At minimum, a functional demonstration of generated scenarios being used by a driving system would help establish executability and initial utility beyond trajectory visualization.

  3. Since CARLA/Chrono already generate the post-impact reference trajectories, what advantages does the learned post-collision generator offer over generating pre-collision trajectories and simulating the subsequent motion?

[1] Feng et al. Trafficgen: Learning to generate diverse and realistic traffic scenarios. ICRA, 2023.

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (part 1/3)

We sincerely appreciate your insightful and constructive comments. Please let us know if we misunderstood any of your questions. We would be happy to provide corrected responses upon your clarification.

W1: The central physical-plausibility claim lacks direct quantitative validation. Evaluation primarily measures displacement and relative-distance errors, which do not establish collision success, absence of interpenetration, or physically consistent impact responses.

Q1: Could the authors directly evaluate the physical validity of the generated trajectories, including collision success rates, interpenetration, and post-impact momentum or impulse consistency?

Thank you for the suggestion. We report the validity of the following metrics.

  1. Collision rate. This measures how frequently the generated outputs produce collisions. A collision is detected if the two vehicles make contact at least once during the prediction horizon. A vehicle is represented by an oriented rectangle of dimension 4.5 m×1.8 m4.5\,\mathrm{m}\times1.8\,\mathrm{m}.

  2. Absence of interpenetration. For this evaluation, we place a 0.9 m×0.9 m0.9\,\mathrm{m}\times0.9\,\mathrm{m} box at the center of each vehicle, aligned with its direction of travel, and define interpenetration as contact or overlap between these two boxes. The interpenetration cases are counted only from the collision samples. That is, we consider the rate of interpenetration events conditional on the collision events. We report the rate of absence of interpenetration.

  3. ΔV\Delta V. This measures the prediction accuracy of the change in the velocity magnitude after collision.

  4. Normal ΔV\Delta V. This measures the momentum consistency as the change in the normal components of velocity after impact. This is associated with the impact response from restitution.

  5. Tangential ΔV\Delta V. This measures the momentum consistency as the change in the tangential components of velocity after impact, which is associated with the impact response from contact friction.

  6. Yaw rate. This measures the rotational validity after impact. The yaw rate is one of our physical scoring metrics (defined Appendix A.5, Eqs. (25)–(27)).

Additional quantitative evaluation of collision occurrence, velocity-change accuracy, absence of interpenetration, physical consistency of impact responses, and rotational validity. Velocity errors and residuals are reported in m/s\mathrm{m/s}, and yaw-rate errors in rad/s\mathrm{rad/s}.

MethodCollision Occurrence (%) ↑\uparrowΔV\Delta V MAE (m/s) ↓\downarrowAbsence of interpenetration (%) ↑\uparrowNormal ΔV\Delta V (m/s) ↓\downarrowTangential ΔV\Delta V (m/s) ↓\downarrowYaw-rate MAE (rad/s) ↓\downarrow
LaneGCN76.575.59560.166.6061.4540.361
HiVT75.675.58662.495.6591.3290.410
HPNet62.515.50166.215.6671.2930.439
TrafficGen64.134.95561.527.2311.9090.612
LCTGen69.655.58559.656.2771.3960.486
InteractTraj68.494.44666.149.2772.9821.623
KING76.305.15093.744.3642.4200.471
COLLIGEN–HiVT94.552.68584.301.9500.7090.316
COLLIGEN–QCNet95.682.12789.852.5840.8070.258

Table shows the results of the additional evaluation. We observe that COLLIGEN outperforms the baselines across all metrics. One exception is the absence of penetration;in that case, KING performs the best. This is because of the property of the KING method. As shown in the qualitative examples in Fig. 4 and Appendix A.11, vehicles simply come to a halt after a collision in KING. Since the vehicles remain still, the likelihood of penetration also decreases, leading to a high rate of absence-of-interpenetration. However, such vehicle behavior under KING is not necessarily optimal, which can be observed from the performance in other metrics.

REVISION ACTION. We added the results to the main text as Table 2 in the revised paper.

AuthorsOpenReview ↗

Rebuttal by Authors (part 2/3)

W2: The practical utility of the generated data remains unclear. Autonomous-driving training and evaluation motivate the work, but no downstream experiment establishes whether the generated trajectories improve a predictor or driving policy, or provide effective safety tests. Reconstruction accuracy alone does not substantiate these intended uses.

Q2: Could the authors demonstrate the utility of the generated data through a focused downstream training or safety-testing experiment, analogous to the policy-training evaluation in TrafficGen [1]? At minimum, a functional demonstration of generated scenarios being used by a driving system would help establish executability and initial utility beyond trajectory visualization.

Downstream training: collision prediction. We thank the reviewer for the insightful suggestion. To investigate the utility of collision data generated by COLLIGEN for downstream training, we consider the target task of collision prediction as follows.

  • A trajectory model is input with trajectories of two nearby vehicles (not yet collided).

  • The model simulates future trajectories of the input: it predicts a collision if the simulation outcome incurs a collision; otherwise, it predicts the input as normal.

  • Two models are compared: one trained with both normal and collision data generated by COLLIGEN; the other with normal trajectories only.

The key ability to test is whether a model can predict future trajectories that lead to collisions. Possible applications include downsizing trajectory generation models (e.g., quantization) or performing knowledge distillation (KD) to student models. Such small models can be embedded in vehicles to support real-time prediction of short-period future trajectories.

Setup. We conducted an experiment of fine-tuning TrafficGen [1] with collision data, where TrafficGen was initially pretrained on normal driving data. Specifically, we partitioned our generated collision data into training, validation, and test sets at the scenario level, using a split ratio of 0.75:0.20:0.05, respectively. We fine-tuned TrafficGen on the training set, selected a checkpoint using the validation set, and evaluated the pretrained and fine-tuned models on the same test set under identical input conditions.

Next, we compared the original and fine-tuned TrafficGen models. Specifically, we conducted inference on 200 scenarios, comprising 100 normal driving scenarios from Argoverse and 100 collision scenarios from reconstructed ground-truth (GT) data. For the collision scenarios, we provided observations from −3-3 to −1-1 seconds before the collision as input and evaluated whether each model could predict the impending collision.

Confusion matrices for collision prediction before and after fine-tuning TrafficGen.

TrafficGen

Actual \ PredictedCollisionNormal
Collision6139
Normal0100

Fine-tuned TrafficGen

Actual \ PredictedCollisionNormal
Collision955
Normal298

Results: a significant drop in missed collision detection. Table (TrafficGen confusion matrices) shows the confusion matrix of the collision prediction. The original TrafficGen failed to predict the collision in 39 out of 100 collision scenarios, whereas the fine-tuned TrafficGen missed only 5 cases. These missed detection, or false-negative (FN) cases, are critical because the model failed to predict collision. The reduction in FN ratio is from 39% to 5% thanks to the fine-tuning with the collision data generated by Colligen.

The collision fine-tuning incurs a slight increase in false alarms, or false positive (FP) cases, where a model predicts a collision from otherwise normal trajectories. Specifically, the base model did not predict any collision or 0% FP rate. For the fine-tuned model, there is only a slight increase of the FP rate to 2%. However, not only the rate is quite small, but also the FP cases can be regarded as less critical errors than FN cases in some scenarios.

Conclusion. Our experiments show that fine-tuning TrafficGen with collision data substantially improves its ability to anticipate impending collisions while maintaining high prediction accuracy in normal driving scenarios. This further demonstrates the practical utility of collision data generated by COLLIGEN for downstream collision prediction tasks, beyond improvements in trajectory reconstruction accuracy.

REVISION ACTION. We will add the collision prediction experiments with TrafficGen to Section 3.2.

AuthorsOpenReview ↗

Rebuttal by Authors (part 3/3)

Q3: Since CARLA/Chrono already generate the post-impact reference trajectories, what advantages does the learned post-collision generator offer over generating pre-collision trajectories and simulating the subsequent motion?

Learning-based vs Rule-based Generators. Thank you for the insightful comment. The biggest difference is that CARLA/Chrono are heavily rule-based simulators; whereas we focus on learning-based generators. The status quo is that learning-based generators have just begun to catch up with the CARLA/Chrono output quality, as shown by our work.

Learning-based simulators are inherently lighter, faster, but tend to be inexact. However, we envision that learning-based generators will eventually dominate over complex rule-based counterparts. This has happened in graphics research: for example, simulating liquid, smoke, human clothing, and hair. Recent interesting examples are world models: Google has recently introduced the 4th generation of Genie [1], which renders world scenes based entirely on a data-driven approach without game or physics engines. While not without limitations, the approach is an important direction for physical AI research.

We observed a similar trend even in our work. Colligen combines the parametric knowledge of the base model with fine-tuning guided by simple physics; yet it is able to generate outputs comparable to CARLA/Chrono, which uses a significantly more complex physics engine. Thus, we believe our work can be a meaningful step towards this research direction.

[1] Bruce, Jake, et al. "Genie: Generative interactive environments." Forty-first international conference on machine learning. 2024.

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

This paper improves upon a shortcoming in how motion prediction datasets provide very little supervision for impact and post impact behaviour. This paper uses a combination of a pretrained predictor, self training together with trajectory perturbations and finally a physics selector to choose a best trajectory. The paper first use a pretrained driving trajectory predictor to generate 2 vehicle collision patterns without using any external car crashing training set. For the training phase, they first use a pre-collision phase, then followed by a post-collision self-training phase. It uses a physics based scoring system and an IMH sampling with perturbed quartic Bézier curve. At inference time, the trajectories with the minimum score calculated from the overall penalty score is selected as it is the most plausible trajectory. The evaluation uses 4583 conditioned trajectories reconstructed with 2 simulators Carla and Chrono, which is validated using correlation analysis with CISS dataset. This study has shown that reported displacement and pairwise distance errors improve substantially over several prediction and baselines. Qualitative analysis also further shows some impact responses and some failure cases.

Critical strengths
  1. This paper targets a problem that is significant enough and has a sensible implementation. This paper targets the shortage of pre and post impact training trajectories and finds a solution to simulate physical knowledge into driving prediction through generated data.

  2. The research has improved reconstruction of collision trajectories and provides good analysis. They have also conducted a second adapted backbone and conducted some score ablations. They have checked simulator reconstructions against partial real world data, and also explained some failure examples.

Critical weaknesses
  1. The physical basis of why central sampler is chosen is not explained properly and no extra explanation is given. The penalty scores that are clipped mean that energy is bounded, hence the target density for the MCMC is not normalizable. The proposal density q is an endpoint mixture and the author does not explain or show the derivations about how it is used in the sampled path.

  2. This study supports reconstructed trajectories more than actual physical simulations. The simulator predictions that are checked against data only cover 0.3 seconds. A.8 provides the criteria but did not show agreement statistics. The held out schedule tuning details are not explain or specified. Aggregate collision, impact-response, and rotational-validity results are not shown by this study. The comparison methods used also do not show any dynamics after collision.

Questions for authors
  1. What is the implemented Markov state, reference measure and proposal density?
  2. What are the score on reference simulator trajectories?
  3. What are CISS agreement and retention summaries? What is split/tuning method you used?
5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (part 1/4)

We sincerely appreciate your insightful and constructive comments. Please let us know if we misunderstood any of your questions. We would be happy to provide corrected responses upon your clarification.

W1-1: The physical basis of why central sampler is chosen is not explained properly and no extra explanation is given.

Choice of sampler. We chose Independent Metropolis–Hastings (IMH) [1] instead of regular MH. A brief explanation was provided in Sec. 2.2 of the original paper. In the following, we provide a more detailed explanation. The explanation will also be added to in the revision.

The primary reason for the choice is the difficulty of trajectory sampling. An MCMC sampling uses proposal distributions. Sampling from the proposal is essentially a perturbation of the current sample (trajectory) to obtain a new one. Suppose we perform sampling with regular MH. The regular MH uses a proposal density of the form τn∼q(⋅∣τn−1)\tau^n \sim q(\cdot|\tau^{n-1}). If the sampling begins with the initial trajectory τ0\tau^0, τn\tau^n will be a trajectory perturbed sequentially for nn times. Any perturbation to a trajectory will likely distort the natural shape of paths. Thus, sequential perturbations will accumulate distortions, resulting in unnatural trajectories.

In contrast, independent MH uses a proposal of the form τn∼q(⋅∣τ0)\tau^n \sim q(\cdot|\tau^0). We can apply a single perturbation to the reference trajectory τ0\tau^0 to obtain a new sample. Thus, the sampled trajectory will not unnaturally deviate from τ0\tau^0 with a controlled perturbation. This also holds true for our proposed smooth perturbation method utilizing Bézier curves. In summary, IMH is a practically better sampler than regular MH for obtaining natural trajectories.

REVISION ACTION: We have added the explanation requested in line 194-201.

[1] Jun S Liu. Metropolized independent sampling with comparisons to rejection sampling and impor- tance sampling. Statistics and computing, 6(2):113–119, 1996.

W1-2: The penalty scores that are clipped mean that energy is bounded, hence the target density for the MCMC is not normalizable.

A: Thank you for the important comment. We will make a simple fix to the normalization. Importantly, it is fixable without any (practical) impact on our sampling method.

Background. Due to the IMH, a trajectory τ\boldsymbol{\tau} is sampled from a some reference trajectory τ0\boldsymbol{\tau}_0 using proposal distribution qq:

τ∼q(⋅∣τ0)\boldsymbol{\tau} \sim q(\cdot | \boldsymbol{\tau}_0)

For proposal q(⋅)q(\cdot), we use distributions with exponential tails, e.g., Gaussian, Laplace, etc., centered around τ0\boldsymbol{\tau}_0. We obtain τ\boldsymbol{\tau} by slightly perturbing reference τ0\boldsymbol{\tau}_0 with qq.

Thus, it is unlikely that the sampled trajectory τ\boldsymbol{\tau} deviates too much from τ0\boldsymbol{\tau}_0, e.g., the two trajectories are apart by 500 meters. That is, for sufficiently large R>0R>0,

q(τ∣τ0)≈0 if ∥τ−τ0∥>Rq(\boldsymbol{\tau} | \boldsymbol{\tau}_0)\approx 0\text{ if }\|\boldsymbol{\tau}-\boldsymbol{\tau}_0\| > R

The probability of sampling trajectories deviating more than RR vanishes. For example, if qq is Gaussian or Laplace (or some proper mixture),

P(∥τ−τ0∥>R)<exp⁡(−Ω(R))P(\|\boldsymbol{\tau}-\boldsymbol{\tau}_0\| > R) < \exp(-\Omega(R))

where Ω(R)\Omega(R) denotes a function that grows at least as fast as linearly in RR.

Modified target density. We define the unnormalized target density as

\begin{cases} \exp\left(-\beta E(\boldsymbol{\tau})\right), & \boldsymbol{\tau}\in C \\ 0, & \text{otherwise} \end{cases}$$ where $C:=\{\boldsymbol{\tau}\mid\|\boldsymbol{\tau}-\boldsymbol{\tau}_0\|\leq R\}$, and $\beta>0$ is the inverse temperature parameter. It can be trivially shown that $\pi$ is normalizable for any finite $R$. The exact value of $R$ is of no practical importance, as long as we set it sufficiently large, as explained below. **No practical impact on sampling.** For theoretical completeness, we will state the modified $\pi$ in the revision. However, importantly, *this has no practical impact on our previously sampled data.* This is because we can assume that $R$ has been set sufficiently large from the beginning, and almost all the data have been sampled from $C$. Thus, the sampled data for our fine-tuning remain unchanged for practical purposes. **REVISION ACTION:** We state the modified target density in Eq. (2). We also explained its implications in lines 162-169.
AuthorsOpenReview ↗

Rebuttal by Authors (part 2/4)

W1-3. The proposal density q is an endpoint mixture and the author does not explain or show the derivations about how it is used in the sampled path.

To help understand our method, we present a simplified toy model.

Toy example. A trajectory is a sequence of points or vehicle locations. For simplicity, assume (i) locations are 1-D points; (ii) a trajectory consists of 3 points. Consider trajectory τ=(xs,xm,xe)\boldsymbol{\tau} = (x_s, x_m, x_e) which has starting point xsx_s, middle point xmx_m, and endpoint xex_e. Suppose we perturb τ\boldsymbol{\tau} with some noise Δ\Delta and obtain τ~\boldsymbol{\tilde\tau} as follows.

τ~=(xs,xm+0.5Δ,xe+Δ)\boldsymbol{\tilde\tau} = (x_s, x_m+0.5 \Delta, x_e+\Delta)

where the starting point is unchanged; the endpoint is perturbed by Δ\Delta; the middle point is perturbed by half of Δ\Delta. Thus, τ\boldsymbol{\tau} and τ~\boldsymbol{\tilde\tau} share the starting point but differ in other waypoints.

Importantly, the perturbation is fully determined by Δ\Delta. Suppose Δ\Delta is sampled from the standard normal distribution. The likelihood of moving from τ\boldsymbol{\tau} to τ~\boldsymbol{\tilde\tau} is

q(τ~∣τ)=N(x~e;xe,1)=N(xe−Δ;xe,1)∝exp⁡(−Δ2/2)q(\boldsymbol{\tilde\tau}| \boldsymbol{\tau}) = \mathcal{N}(\tilde x_e; x_e,1) = \mathcal{N}(x_e-\Delta;x_e,1) \propto \exp(-\Delta^2/2)

where N(⋅;μ,σ2)\mathcal{N}(\cdot;\mu,\sigma^2) denotes the Gaussian density. The perturbation density q(τ~∣τ)q(\boldsymbol{\tilde\tau}| \boldsymbol{\tau}) is fully characterized by the displacement at the endpoint, or Δ\Delta.

Actual perturbation. The perturbation in our paper can be understood by extending the toy example as follows.

  • A trajectory is a sequence of 2D location points.
  • Trajectories are reconstructed as Bézier curves, which are fully specified by control points.
  • The control points are perturbed by random Δ∈R2\Delta\in\mathbb{R}^2. The last control point, or the endpoint, is perturbed by Δ\Delta. The intermediate points are perturbed by less than Δ\Delta to make the perturbation gradual and smooth.

We use the 2D-Laplace distribution to sample Δ\Delta. Similar to the toy example, the perturbation is fully determined by endpoint displacement Δ\Delta.

Proposal density with HiVT. The base model HiVT [1] is used in our paper. The proposal density mixes HiVT sampling with perturbation. HiVT outputs K=6K=6 plausible variations τ1,…,τ6\boldsymbol{\tau}_1,\dots,\boldsymbol{\tau}_6 of the base trajectory τ0\boldsymbol{\tau}_0, each with probability pkp_k. We perturb each of the output trajectories with Laplace noise as explained above. This leads to a mixture distribution of outputs. In summary,

  • Sample a trajectory from τ1,…,τ6\boldsymbol{\tau}_1,\dots,\boldsymbol{\tau}_6 with probability p1,…,p6p_1,\dots,p_6 from HiVT.
  • Perturb the sampled τk\boldsymbol{\tau}_k to obtain τ~\boldsymbol{\tilde\tau}. This perturbation has the likelihood Laplace(τ~(T);τk(T),bk)\text{Laplace}(\boldsymbol{\tilde\tau}(T); \boldsymbol{\tau}_k(T),\mathbf{b}_k) defined by endpoints.

Thus, the overall likelihood is the mixture

q(\boldsymbol{\tilde\tau} | \boldsymbol{\tau}_0) &=\sum_{k=1}^6 P(\boldsymbol{\tau}_k|\boldsymbol{\tau}_0)q(\boldsymbol{\tilde\tau} | \boldsymbol{\tau}_k,\boldsymbol{\tau}_0)\\\\ &=\sum_{k=1}^6p_k\cdot\text{Laplace}(\boldsymbol{\tilde\tau}(T); \boldsymbol{\tau}_k(T),\mathbf{b}_k) \end{aligned}$$ which is our proposal density. **REVISION ACTION:** We added the explanation of why the proposal can be defined as an endpoint mixture in Eq. (4). [1] Zhou, Zikang, et al. "Hivt: Hierarchical vector transformer for multi-agent motion prediction." 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022.
AuthorsOpenReview ↗

Rebuttal by Authors (part 3/4)

Q1: What is the implemented Markov state, reference measure and proposal density?

A. We summarize the discussion on the sampling method as follows.

  • Markov state: The vehicle trajectory τ\boldsymbol{\tau}, a time sequence of 2D vehicle locations.
  • Reference measure: The target distribution is π(τ)\pi(\boldsymbol{\tau}), a Boltzmann-like density exp⁡(−βE(τ))\exp(-\beta E(\boldsymbol{\tau})) restricted to C={τ∣∥τ−τ0∥≤R}C=\{\boldsymbol{\tau}\mid\|\boldsymbol{\tau}- \boldsymbol{\tau}_0\|\leq R\} for some sufficiently large RR.
  • Proposal density: For the HiVT base model, the proposal density is a mixture of Laplace distributions of trajectory endpoints.

W2-1: This study supports reconstructed trajectories more than actual physical simulations. The simulator predictions that are checked against data only cover 0.3 seconds.

Evaluation duration.

The physical data from NHTSA is available for the following 0.6 second durations: 0.3 seconds pre-collision and 0.3 seconds post-collision (diagram in Fig. 7 of the paper). Our simulator predictions are validated for this 0.6 second duration (please refer to Appendix A.8.1 line 1023).

In our original manuscript, we stated in line 969 that we validated 0.3 second duration of post-collision. This might have been confusing: we actually validated both pre- and post-collision durations (each 0.3 seconds and 0.6 seconds total). In A.8.2, we validated both pre- and post-collision durations (as stated in line 1023). We have corrected the statement in the revision.

Physical data availability.

To our knowledge, this NHTSA data is the only available physical data on vehicle collisions. The limited availability of data challenges collision research. This is why we used two simulators, CARLA and Chrono, to validate simulations. We would be grateful if you could understand these difficulties.

REVISION ACTION. We corrected our statement in Appendix A.8.1 that the simulation is validated for 0.6 seconds encompassing pre- and post-collision durations.

W2-2: A.8 provides the criteria but did not show agreement statistics.

Q3-1: What are CISS agreement and retention summaries?

CISS agreement.

We analyze the CISS agreement for pre-collison and post-collision data. For pre-collision agreement, we evaluate the vehicle speed; for post-collision agreement, we evaluate the change in ΔV\Delta V. To assess reconstruction fidelity, we computed the agreement statistics across 375,240 matched scalar observation points. Each matched point pairs an observed and reconstructed value for the same vehicle, event, component, and observation time.

Table 9 and 10 show the agreement statistics for the retained simulation by CARLA/Chrono. We observe that the level of agreement is significantly high for both pre- and post-collision statistics.

Table: Pre-collision agreement statistics of Speed for retained CARLA/Chrono reconstructions.

MetricValue
MAE (km/h\mathrm{km/h})0.0026
Bias (km/h\mathrm{km/h})-0.0002
Nominal 95% limits of agreement (km/h\mathrm{km/h})[−0.0784, 0.0779][-0.0784,\,0.0779]
Pearson rr0.9999
OLS slope0.9999
OLS intercept (km/h\mathrm{km/h})0.0012

Table: Post-collision agreement statistics of ΔV\Delta V for retained CARLA/Chrono reconstructions.

MetricValue
MAE (km/h\mathrm{km/h})0.0973
Bias (km/h\mathrm{km/h})-0.0105
Nominal 95% limits of agreement (km/h\mathrm{km/h})[−0.3676, 0.3467][-0.3676,\,0.3467]
Pearson rr0.9999
OLS slope0.9981
OLS intercept (km/h\mathrm{km/h})-0.0222

Retention summaries.

From NHTSA data, we extracted 9,349 cases of two-vehicle collisions. Among those cases, there were 5,067 cases that had sufficient CISS data for trajectory reconstruction with CARLA/Chrono. Among those reconstructed cases, 4,583 cases passed the criteria for retention, i.e., the simulated results matched well with the NHTSA pre- and post-collision data. Thus, the retention rate was 4,583/5,067=90.45%4,583/5,067=90.45\%.

REVISION ACTION: We added the CISS agreement and retention summaries in Appendix A.8.2.

W2-2. The held out schedule tuning details are not explain or specified.

Q3-2: What is split/tuning method you used?

Our fine-tuning at each stage uses a mixture of normal-driving and pseudo-collision data (Appendix A.2). The normal-to-collision data ratio is scheduled as 4:1, 4:2, and 4:3 over three stages. At each fine-tuning stage, the normal and collision data are newly split into 90% training and 10% validation. The split is applied independently to normal and collision data. The training split of normal and collision data is merged to form the training data for the current stage. The same applies to the validation data. The tuning of the model follows the training procedure of the base model, e.g., HiVT.

REVISION ACTION: We added the explanation of split/tuning method in line 741 of Appendix. A.2.

Reviewer_4UMUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper here proposes a method to adapt pretrained normal-driving trajectory predictors into generators of two-vehicle pre- and post-collision trajectories without requiring supervised crash data during training. The authors introduce an Independent sampler that perturbs control points and evaluates candidates using energy functions. This method is then evaluated over an author constructed benchmark cross validating NHSTA records.

Critical strengths
  • Formulation is an original and well-motivated approach to adapting nominal motion predictors into a joint pre/post motion prediction / generation without need for external simulator.
  • The proposed traj proposal mechanism is a substantial speedup against traditional method of MC dropout or other perturbation methods.
  • New evaluation benchmark that simulates nhsta data base is novel and provides valuable resource for evaluating safety critical traj generation.
Critical weaknesses
  • The formulation between the two phase is confusing and maybe contradictory between sections and appendix.
    • Section 2.1 states that total time between pre and post collision phase is T (3s?) but then in Algo 1 and appendix A.5 authors states that pre phase is a full 3s trajectory and then phase 2 (theta_post) takes input 2s upto t* and then regenerates 3s trajectory ?
    • t* is defined as the point where both vehicles are closest to each other, how is that different from collision and if that is collision, what does line 256 mean when it says delta_t_pre is remaining time of pre-collision phase?
  • There are several reproducibility gaps across the experimental tables and protocols, including unexplained numerical discrepancies between the main results in Tables 1–2 and the identical default schedules in Tables 4–5 and Figure 5, asymmetric ground-truth conditioning?
  • Eq 7 does not explain the event when two vehicle collides and continues in opposite direction. In this case mu_ground can not be applied to the total of mass?
  • T notion is overused (one for temperature and one for time horizon.
  • The evaluation section is missing baselines that computes pre and post collision trajectories from nhtsa reports. Consider AccidentSim
Questions for authors

Can you clarify the exact algorithm on what is the length of trajectories in pre and post collision and in phase 1 and phase 2. Very specifically, what is the length of trajectory input to the predictor and output of that predictor for phase x pre/post collision. And is it consistent with text in paper and in appendix?

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (part 1/4)

We sincerely appreciate your insightful and constructive comments. Please let us know if we misunderstood any of your questions. We would be happy to provide corrected responses upon your clarification.

W1: The formulation between the two phases is confusing and maybe contradictory between sections and appendix.

Section 2.1 states that total time between pre- and post-collision phases is TT (3s?), but then in Algorithm 1 and Appendix A.5, the authors state that the pre-phase is a full 3s trajectory and then phase 2 (θpost\theta_{\text{post}}) takes input 2s up to t* and then regenerates a 3s trajectory?

W1-2: t* is defined as the point where both vehicles are closest to each other. How is that different from collision? And if that is collision, what does line 256 mean when it says Δtpre\Delta t_{\text{pre}} is the remaining time of the pre-collision phase?

Q1: Can you clarify the exact algorithm on what is the length of trajectories in pre and post collision and in phase 1 and phase 2. Very specifically, what is the length of trajectory input to the predictor and output of that predictor for phase x pre/post collision. And is it consistent with text in paper and in appendix?

Thank you for the question. Building on the distinction between the timelines explained in our response to W1, we clarify below the lengths of the trajectories provided to and generated by the predictor at each stage. The sampling interval is 0.1s0.1\mathrm{s}, and both training phases use a 20-step input and a 30-step output.

StageInputOutput
Phase 1Original history: [-3,-1) 20 steps (2s)Future trajectory: [-1, 2) 30 steps (3 s) include t_c
Phase 2[t_c-20, t_c)(t_c, t_c+30]
Final inferenceOriginal history: 20 steps (2s)Pre + post trajectory: 30 steps (3s)

The input to Phase 2 consists of the most recent 20 steps, including the collision frame, namely the samples from tc−1.9st_c-1.9\mathrm{s} to tct_c. Its output consists of 30 new steps from tc+0.1st_c+0.1\mathrm{s} to tc+3.0st_c+3.0\mathrm{s}. In the concatenated trajectory used during training, the Phase 1 trajectory is retained up to and including the collision frame, and the subsequent segment is replaced by the Phase 2 output. Therefore, the length of this concatenated trajectory can exceed 30 steps.

At final inference, the model generates a new 30-step trajectory containing pre- and post-collision segments from the original 20-step input. Within this final 30-step trajectory, the lengths of the pre- and post-collision segments vary with the generated collision time, while their combined length remains 30 steps.

These stage-specific input and output configurations correspond to the settings described in Algorithm 1 and Appendix A.5. Section 2.1 was intended to describe the division of the final prediction horizon into pre- and post-collision segments, but we agree that this can be confused with the distinction between the training phases. We will use distinct notation for the durations of the segments within the final prediction horizon and explicitly specify the 20-step input, 30-step output, and concatenation procedure for each phase in both the main text and the appendix to ensure a consistent explanation.

AuthorsOpenReview ↗

Rebuttal by Authors (part 2/4)

For line 256 : The closest-approach time t⋆t^\star is not necessarily a collision time: it is the time at which the distance between the two vehicles is smallest within the generated trajectory. The vehicles can remain separated even at this time. An actual collision is determined separately by checking contact or overlap between their geometries.

The purpose of Eq. (7) is to provide a distance-based reachability penalty, rather than to identify an actual collision. To clarify line 256, we will explicitly define the remaining time as

Δtremain=max⁡(0,T−t⋆)\Delta t_{\mathrm{remain}} = \max(0,T-t^\star)

where TT denotes the duration of the prediction horizon used for pre-collision generation. Thus, "remaining time" refers to the portion of this horizon remaining at t⋆t^\star, not the time until an actual collision occurs.

For example, consider a 30-step prediction horizon with Δt=0.1 s\Delta t=0.1\,\mathrm{s} and two candidates whose closest-approach states occur at step 25. Five steps, or 0.5 s0.5\,\mathrm{s}, remain within the horizon. Under the speed-based approximation in Eq. (7), suppose the first candidate's separation would require ten additional steps to close, whereas the second candidate's separation could be closed within five steps. The first candidate exceeds the reachability bound over the remaining horizon and therefore has a positive distance residual. The second satisfies this bound and has zero distance residual.

This comparison concerns an optimistic reachability bound based on the vehicles' speeds; it does not assert that the generated continuations actually collide. When incorporated into the physics-guided sampling score, the residual discourages candidates whose separation is too large to close within the remaining horizon. Through subsequent fine-tuning on the generated samples, this is intended to encourage collision generation within the three-second prediction horizon. However, the residual alone does not guarantee a collision, so collision occurrence and timing are verified separately using the collision detector.

We will revise line 256 to state "the remaining duration of the pre-collision prediction horizon evaluated at t⋆t^\star" and clarify the distinction between closest approach, the reachability penalty, and actual collision detection.

REVISION ACTION: We clarified the annotation of Δtpre\Delta t_{\mathrm{pre}} and Δtpost\Delta t_{\mathrm{post}}. We moved the description originally at line 256 to Sec. 2.4.

AuthorsOpenReview ↗

Rebuttal by Authors (part 3/4)

W2: There are several reproducibility gaps across the experimental tables and protocols, including unexplained numerical discrepancies between the main results in Tables 1–2 and the identical default schedules in Tables 4–5 and Figure 5, asymmetric ground-truth conditioning?

We thank the reviewer for pointing out the issue. There were different experimental settings and evaluation metrics used in Tables 1–2, Tables 4–5, and Figure 5, which were not sufficiently explained. We will explain the metrics and settings as follows.

Table 1 and 3 compare COLLIGEN with baseline methods and evaluate the contributions of its physics-based scoring functions. Some of the baselines require road map information to generate trajectories. Thus, for a fair comparison, our method also used map information to obtain results in Table 1-2. Adding the map information has some regularization effect on the vehicle trajectory, i.e., it softly aligns the trajectory to pre-defined road lanes irrespective of the vehicles' expected physical behavior such as post-impact dynamics. This makes the error performance slightly worse than the inference without map information.

Tables 4–5 evaluate the effects of the data-merge schedule α\alpha and physical constraint schedule γ\gamma, respectively. This is ablation study, i.e., self-comparison. Thus we did not need the map information, and thus the map was not input to the model during inference. This leads slightly better performance (lower ADE, FDEs) than the baseline comparison. Also, we considered that incorporating map information could influence the evaluation of the effects of α\alpha and γ\gamma.

For the sake of consistency, we conducted additional experiments for Tables 4–5 with map information included in both training and inference. The updated results are presented below, using the same map-conditioned setting as Tables 1–2.

Comparison of the stage-wise physical constraint schedules α\alpha with map information. Loose and Strict use α=(0.25,0.50,0.75)\alpha=(0.25,0.50,0.75).

α(1),α(2),α(3)\alpha^{(1)}, \alpha^{(2)}, \alpha^{(3)}PreADE ↓\downarrowPreFDE ↓\downarrowPostADE ↓\downarrowPostFDE ↓\downarrowrel_dist_err ↓\downarrow
1.00,1.00,1.001.00, 1.00, 1.00 (Const.)2.3092.2411.9732.9932.519
0.25,0.50,0.75\mathbf{0.25, 0.50, 0.75} (Ours)1.7061.5931.4872.1382.007
0.75,0.50,0.250.75, 0.50, 0.25 (Aggr.)1.9751.8381.6582.3282.210

Comparison of the stage-wise data merge schedules γ\gamma with map information. Const. and Aggr. use γ=(1.0,0.9,0.8)\gamma=(1.0,0.9,0.8).

γ(1),γ(2),γ(3)\gamma^{(1)}, \gamma^{(2)}, \gamma^{(3)}PreADE ↓\downarrowPreFDE ↓\downarrowPostADE ↓\downarrowPostFDE ↓\downarrowrel_dist_err ↓\downarrow
1.2, 1.0, 0.81.2,\ 1.0,\ 0.8 (Loose)2.5832.4922.0773.1292.572
1.0, 0.9, 0.8\mathbf{1.0,\ 0.9,\ 0.8} (Ours)1.7061.5931.4872.1382.007
1.0, 0.8, 0.61.0,\ 0.8,\ 0.6 (Strict)1.7801.6701.4872.0012.079

Figure 5 uses metrics FullADE and FullFDE whose definitions were missing. FullADE is a step-weighted average of the corresponding PreADE and PostADE values for each scenario. FullFDE corresponds to PostFDE when the post-collision segment includes the final prediction step.

REVISION ACTION: We revised Table 4–5 to present results on inference with map information.

AuthorsOpenReview ↗

Rebuttal by Authors (part 4/4)

W3: Eq. (7) does not explain the event when two vehicles collide and continue in opposite directions. In this case, μground\mu_{\text{ground}} cannot be applied to the total mass?

We thank the reviewer for raising this point. The expression in the original paper described a special case for simplicity. We will provide the expression for the general case, where our implementation follows this general case. The coefficient μgnd\mu_{\mathrm{gnd}} represents the shared road condition. The magnitude and direction of the frictional impulse are determined separately for each vehicle as follows.

Let pi−\mathbf{p}_i^- denote the momentum of vehicle ii from the previous time step for i=1,2i=1,2. Under our sliding-friction approximation, the ground-friction impulse at vehicle ii is

Jfric,i=−min⁡!(μgndmigΔt,,∥pi−∥2)ui−,i∈1,2\mathbf{J}_{\mathrm{fric},i} = -\min\\!\left(\mu_{\mathrm{gnd}}m_i g\Delta t,\\, \lVert\mathbf{p}_{i}^{-}\rVert_2\right)\mathbf{u}_{i}^{-}, \qquad i\in\\{1,2\\}

with ui−\mathbf{u}_{i}^{-} is the unit vector in the direction of pi−\mathbf{p}_{i}^{-}. Jfric,i\mathbf{J}_{\mathrm{fric},i} has the opposite direction to the original momentum of vehicle ii and can be considered as the loss of momentum due to ground friction. The expected momentum is the sum of the original momentum and the ground-friction impulse:

pexp+=p−+Jfric\mathbf{p}_{\mathrm{exp}}^{+} = \mathbf{p}^{-} + \mathbf{J}_{\mathrm{fric}}

where Jfric=∑i=12Jfric,i\mathbf{J}_{\mathrm{fric}} = \sum_{i=1}^{2}\mathbf{J}_{\mathrm{fric},i} and p−=∑i=12pi−\mathbf{p}^- = \sum_{i=1}^{2}\mathbf{p}_i^-. Thus, when the vehicles move in opposite directions, their frictional impulses can partially cancel the original total momentum.

REVISION ACTION: We provided the expression in line 277-286 of Sec. 2.3.


W4: T notion is overused (one for temperature and one for time horizon.

Thank you for the comment. We defined a new parameter β:=T−1\beta := \mathcal{T}^{-1}, or the inverse temperature. Thus, the target density is defined as exp⁡(−βE)\exp(-\beta E) instead of exp⁡(−ET)\exp(-\frac{E}{\mathcal{T}}).


W5: The evaluation section is missing baselines that computes pre and post collision trajectories from nhtsa reports. Consider AccidentSim.

We thank the reviewer for this suggestion. We examined methods that generate collision trajectories from NHTSA reports, including AccidentSim[1], TRACE[2], and AC3R[3].

AccidentSim is particularly relevant because it combines pre-collision trajectory planning with a learned post-collision predictor, AccidentLLM. However, this work does not provide a publicly available official implementation or pretrained checkpoint, which makes the comparison difficult.

TRACE and AC3R are simulator-based accident reconstruction pipelines. TRACE estimates vehicle states and constructs paths from crash reports and road information before executing the scenarios in CARLA. AC3R extracts information from police reports and constructs collision paths for execution in BeamNG. Although these methods reconstruct accidents, their post-collision motion is computed by the underlying physics simulators. Thus, they were not directly comparable to Colligen.

We were not able to find further baselines; still we would welcome any suggestions for additional publicly available methods that generate both pre- and post-collision trajectories and support evaluation under comparable input conditions.

[1] Zhang, Xiangwen, et al. "Accidentsim: Generating physically realistic vehicle collision videos from real-world accident reports." arXiv e-prints (2025): arXiv-2503.

[2] Salsabil, Nahian, and Sebastian Elbaum. "TRACE: Topology-aware Reconstruction of Accidents in CARLA for AV Evaluation." Proceedings of the 34th ACM International Conference on the Foundations of Software Engineering. 2026.

[3] Huynh, Tri, Alessio Gambi, and Gordon Fraser. "AC3R: automatically reconstructing car crashes from police reports." 2019 IEEE/ACM 41st International Conference on Software Engineering: Companion Proceedings (ICSE-Companion). IEEE, 2019.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.