Scenario Diffusion: Controllable Driving Scenario Generation With Diffusion

Automated creation of synthetic traffic scenarios is a key part of validating the safety of autonomous vehicles (AVs). In this paper, we propose Scenario Diffusion, a novel diffusion-based architecture for generating traffic scenarios that enables controllable scenario generation. We combine latent diffusion, object detection and trajectory regression to generate distributions of synthetic agent poses, orientations and trajectories simultaneously. To provide additional control over the generated scenario, this distribution is conditioned on a map and sets of tokens describing the desired scenario. We show that our approach has sufficient expressive capacity to model diverse traffic patterns and generalizes to different geographical regions.

Paper

References (40)

Scroll for more · 28 remaining

Similar papers

Peer review

Reviewer 9X4d4/10 · confidence 5/52023-06-21

Summary

This work adopts a conditional diffusion model to generate the abstract driving scene.

Strengths

It has great potentials to generate future scenes with diffusion model.

Weaknesses

(W1) No comparsion with existing works at all. It is significant to compare with previous methods to demonstrate the superiorty of the proposed method in terms of diversity, generalizability or other aspects. Quantitative comparisons with methods such as SceneGen [1], TrafficGen [2], CTG [3], and Scene Diffusion [4] would help to make the improvement more convincing and easier to recognize. Could the authors provide related results or briefly discuss why these methods were not included in the comparison? (W2) The proposed method aims to generate all agents simultaneously so that the relationships between them can be learned. Nevertheless, auto-regressive manner also embodies such functionality as it can condition on the generated agents. Have the authors explored the use of auto-regressive ways within the proposed framework? (W3) As a generation work, a great potential would be producing additional data for training, customizing safety critical scenes for evaluation. However, this part is missing from the current version, which means the claimed benefit is unjustified at all. Could the authors complement some experiments to support that? [1] SceneGen: Learning to Generate Realistic Traffic Scenes [2] TrafficGen: Learning to Generate Diverse and Realistic Traffic Scenarios [3] Guided Conditional Diffusion for Controllable Traffic Simulation [4] Generating Driving Scenes with Diffusion

Questions

See weakness for details. I am mostly concerned that this work only decribe their method but do not compare with rich existing literatures at all. It is hard to tell whether the proposed method really works and is better than existing works.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

1 poor

Presentation

2 fair

Contribution

2 fair

Limitations

No concern.

Reviewer VGNk6/10 · confidence 4/52023-06-26

Summary

This paper proposes a novel approach to use latent diffusion for driving scenario generation. The generated driving scenario is defined as bounding boxes with associated trajectories. The proposed diffusion model is comprised of an autoencoder to encode the BEV image into latent space and a conditional UNet for diffusion modeling. To increase the controllability of the generated scenario, map conditioning and token conditioning are incorporated into the process. The experiments are conducted both on the large-scale internal dataset and the Argoverse 2 dataset, which validates the effectiveness of the proposed algorithm and its flexibility for controllable generation.

Strengths

1. The idea of applying latent diffusion to driving scenario generation is novel and interesting, as diffusion methods have shown great capability in generative tasks. 2. The proposed token conditioning mechanism brings more controllability and the possibility of customization for this task. 4. The overall writing and presentation of the paper are clear and easy to follow.

Weaknesses

1. For the metrics to evaluate the generated scenarios, it is very important to evaluate the validity of the driving scenario, but the paper only measures the rate of off-road trajectories. Other aspects like the rate of collisions between agents, smoothness and plausibility of the trajectories, and possible traffic rules violations should also be measured. 2. Although in Sec 4.3.2 some examples have shown that the proposed method is aware of the relationship between agents, using BEV encoded image to represent the scene without specific designs for agents interactions could not well capture the relationships between agents compared with those GNN-based methods shown in motion-forecasting. Having specific designs to explicitly model the possible interactions is important for the complex driving scenario generation. 4. There is no direct quantitative measurement of how the generated sample faithfully follows the conditioning agent and global token. 3. The goal of scenario generation is to facilitate the development of other autonomous driving tasks like motion forecasting or planning. Therefore, a better way to show the effectiveness and diversity of the generated scenario is by testing them on downstream tasks. Some simple experiments showing the performances of a motion forecasting model training on real and generated data would greatly improve the value of the proposed method.

Questions

1. During the sampling process, would it be helpful to use techniques similar to classifier-free guidance to make the sample ground on the conditioning token more faithfully? 2. Currently the agent token still needs very specific values of different aspects to define its behavior, is it possible to use some more abstract behavior tokens to only specify its overall behavior like overtaking, speed up, right turn, etc. to promote better diversity without much engineering effort?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have well discussed the limitations and the potential negative societal impact.

Reviewer uQ2d4/10 · confidence 4/52023-07-04

Summary

This paper tackles the problem of generating both the initial configuration and vehicle trajectories for driving scenarios given an input top-down map. It proposes a latent diffusion model that operates in the latent space of a learned autoencoder trained to decode vehicle boxes and trajectories. The model can be directly conditioned to encourage a subset of agents to adhere to user specifications. Experiments show that once trained on a large real-world driving dataset, the model can produce reasonable scenarios at multiple scales and allows for user control over a subset of agents.

Strengths

The problem is important for AV testing and training. The proposed method simultaneously generates both an initial scene configuration and vehicle trajectories, instead of in two stages as in prior work. The proposed latent diffusion model is novel for vehicle trajectories – all prior work I have seen is explicit trajectory diffusion and the design of the autoencoder for such a model is not trivial. The model allows directly (optionally) conditioning on attributes of one or more agents which gives nice flexibility and allows generating variations of scenarios directly taken from e.g. a log or specified by a user. Results of generating short 4 sec scenarios look reasonable. Experiments show the ability to generate scenarios of various sizes on the sizable Argoverse 2 dataset. Fig 4 is a neat demo showing that the model is picking up on the semantics of the relationships between vehicles. The appendix is extensive and provides most of the necessary details to enable reproducibility.

Weaknesses

There are a couple claims that seem over-stated and should be toned down or clarified: * Conditioning tokens are “more abstract features” (L208) that does not require “fully specifying box and trajectory for the agent” (L206) or “having observed such agents in the real world” (L43) – if I understand correctly, these tokens require a user to give the model the desired agent position, heading, length, width, speed, and sometimes curvature over time (L211), which is quite a low-level specification of the trajectory and likely will need to be taken from a log. Or do tokens describe the state only at a single $t=0$? If so, this should be made more clear, and I still would not say this is “abstract”. * The model “generalizes to different geographical regions” (L9) – Table 3 shows seemingly limited ability to generalize across regions with models trained specifically on each region far outperforming other single-region models. The model trained on all regions does reasonably well on each, but this is not generalizing across regions. Moreover, it would be helpful to add a metric that indicates similarity between regions (e.g. MMD between ground truth scenes in A and B) to put the numbers in context. While this experiment is informative about performance across region, it may be a stretch to say it demonstrates generalization across regions. Some limitations of the technical approach may make the proposed method less practically useful: * The agent tokens are not hard constraints and in results like Fig 3 the specified agents move around between samples and don’t exactly meet user specifications. A natural solution for diffusion would be to use some kind of test-time guidance as in [37] and [Trace and Pace, Rempe et al., CVPR 2023]. * The 4 sec time horizon (2 sec past + 2 sec future) seems quite short and limits the possible scenarios it can capture. * The conditioning requires a user to specify all attributes of an agent. But it seems with the masking strategy it should be possible to condition on partial information (e.g. specify the past trajectory but generate the future, or only specifying speed) which would greatly enhance the flexibility. Some additional evaluation metrics would help to evaluate realism and diversity. For example, MMD on speeds, collision rate between vehicles, and a diversity metric (like average pairwise distance between multiple samples). Additionally, some key experiments are missing: * The accuracy of the model’s controllability is never evaluated, which is the main claim of the paper. It would be good to report how closely generated agents follow the agent tokens specified by the user (e.g. L2 distance between positions, headings, speed, etc..). The 0% model in Table 2 kind of does this, but the reported MMD metric is not very intuitive. * A comparison to a baseline to justify the advantages a diffusion model to generate both static states and dynamic trajectories jointly. E.g. a comparison to TrafficGen [6] or a simpler baseline that only generates static states and then uses a trajectory prediction or traffic simulation model to roll out motion, e.g. [BITS: Bi-level Imitation for Traffic Simulation, Xu et al., ICRA 2023]. Or evaluating if the trajectories generated by the diffusion model align with those from a SOTA trajectory forecasting model? * It would be good to see an example of how these scenarios can be useful. For example, that training a driving policy on a dataset augmented with generated scenarios does better than just a real-world dataset. * More qualitative results (ideally videos, but additional figures in the supplement would also be helpful).

Questions

Overall, I am borderline on this paper. While the latent diffusion formulation is novel and technically sound and the model nicely generates both starting positions and motion for the scenario, the primary contribution of “controllability” is not sufficiently evaluated, and additional evaluations/baselines would strengthen the paper. In the rebuttal, I would really like the authors to clarify my confusion on the agent tokens and ideally evaluate the controllability of the model by showing how closely it follows user specifications. The method Sec 3 at times is rather high level and leaves some questions: * Sec 3.1: why is the encoder not conditioned on the map like in the typical CVAE formulation? * Sec 3.3: A lot of info about tokens are relegated to experiments in Sec 4.3.1 or the appendix but may be better to include in the methods. I’m still left wondering how exactly the cross attention works, e.g. how are the tokens associated with agents in the scene since there is no canonical agent IDs or ordering? Also, if tokens are specifying the properties of agent trajectories, why do multiple samples of the same scenario (e.g. Figs 1, 7, 8) exhibit varying trajectories for the specified agents? * How are the number of agents in the scene determined by the diffusion model? Through the agent probability? Why is the global scene token represented as the masking probability and not directly the number of agents that are desired in the scene? * The methods mentions 2 agent tokens are used, but then 3 are used in Fig 3 and N are used in Table 2, so it’s not clear what is the problem formulation and what’s used in practice. Another other small suggestion: since explicit trajectory diffusion models are becoming more prevalent [37] [Trace and Pace, Rempe et al., CVPR 2023] [Stochastic trajectory prediction via motion indeterminancy diffusion, Gu et al., CVPR 2022] [Planning with Diffusion for Flexible Behavior Synthesis, Janner et al., ICML 2022], it would be good to discuss them in related work and compare to the proposed latent diffusion. ===================== After Rebuttal ====================== After considering other reviews and discussion with the authors and reviewers, I have decided to keep my rating leaning towards reject. Given that the problem is relatively new and the latent diffusion approach is novel, I am willing to be lenient on the lack of a baseline comparison. Especially since TrafficGen (ICRA 2023) is concurrent work and SceneGen (CVPR 2021) did not release code. However, without a baseline, the experiments and metrics need to be comprehensive and convincing to show generated scenes are realistic and match user specifications; I don’t believe this is the case in the current draft. The rebuttal Table 1 adds some new map-based metrics, but important realism evaluations like collision rate, plausibility of trajectories (smoothness and comfort), and diversity are still missing. Moreover, the added evaluation of controllability in rebuttal Figure 3 is rather coarse (considered “correct” if within 2.2 m) and does not consider other user specifications like length, width, speed, and curvature. I also think it would be a strong addition to show results on higher-level semantic conditioning of the model such as vehicle type and maneuver type (e.g. lane change, turning, or u-turn) to justify the claim that the model can handle “abstract” specifications rather than just partial explicit specifications.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

Limitations are sufficiently discussed in Sec 6.

Reviewer cvm44/10 · confidence 4/52023-07-05

Summary

This paper presents a conditional latent diffusion model (LDM) for generating oriented BEV vehicle bounding boxes, each associated with a 4-second trajectory (2s past, 2s future) at a 1s temporal resolution. The autoencoder component of this LDM uses an object detection architecture similar to CenterNet, with rasterized agent inputs. The diffusion model is conditionable on the local map, a global traffic density parameter, and an optional per-agent attribute descriptor. The model is validated on an internal proprietary dataset and Argoverse 2 via a series of ablation studies.

Strengths

What stands out in this work is the overall simplicity and elegance of the framework. It draws on ideas from different communities (generative modeling, object detection, traffic simulation). It also includes interesting original ideas that are potentially applicable to other generative modeling tasks in robotics and self-driving, such as (1) the imbalanced VAE (raster input, vector output); and (2) the masking idea to make agent-level conditioning optional. The writing is clear and mostly easy to follow.

Weaknesses

The key weakness of this work is the lack of any baselines in the experimental analysis. Note that this is not a completely novel task (Section 5 discusses prior techniques). While it is true that not all prior work may be conditionable on per-agent attributes like the proposed model, I believe a comparison to prior work (potentially in the setting without agent token conditioning, as in the last column of Table 2) is warranted to meaningfully understand how well diffusion models perform this task. Secondly, the diffusion model architecture is hard to understand from the main paper in its current form, and a clear description is only available in the supplementary. For example, L_{KL} is marked in Fig. 2a but not described at any point in the main text.

Questions

1. Would it be possible to include one heuristic and one autoregressive approach (as mentioned in Section 5) for the Argoverse dataset as additional columns in Table 2? This is the main limitation of the paper from my perspective, and I am happy to raise my score if it is addressed, or a convincing reason for why such baselines are not useful is provided. 2. Since the conditioning (controlability) is pitched as a key contribution, this should be clearly understandable from Fig. 2b. Would it be possible to expand “M” to show both its encoder and decoder in Fig. 2, with more clarity on how the different inputs/conditions are passed into the diffusion model architecture, instead of a single input arrow? Minor: 1. I couldn’t understand the caption in Table 3, specifically the text enclosed in brackets 2. [34] and [37] have incorrect citation “year” fields (2022 → 2023) Update: Thanks a lot for your clarifications and the efforts made to update the draft. While I no longer have any issues regarding the method/presentation, I believe the rebuttal does not fully address the concern regarding limited baselines in the analysis of the proposed model. The other reviewers have also commented on the insufficiency of the current experiments to demonstrate a high degree of realism and controllability with the proposed model, and I concur. Therefore, I am maintaining my rating of borderline reject.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The key limitations (such as the currently limited generalization capability and simple task definition) are discussed in Section 6.

Reviewer uQ2d2023-08-16

Followup question

Thank you to the authors for the detailed response and clarifications, I really appreciate it. The authors stated that they “have experiments with these higher level descriptions” like “bus” vs “car” tokens instead of the vehicle’s size, and the type of motion (e.g. lane change, turning, or u-turn).” Are these experiments included in the paper or supp? These would be exciting results and showing these examples would greatly strengthen the argument that the model can handle “abstract” specifications. Otherwise, please consider re-phrasing the description to be “partial” specification rather than “abstract”.

Reviewer VGNk2023-08-19

Response to the rebuttal

Thanks to the author for further clarifications. After reading the rebuttal and other reviews, my main concern is the evaluation part in terms of diverse and proper metrics for this task, possible baselines, and more ablations.

Reviewer 9X4d2023-08-20

Response

Thanks for the reviewers' replies. After discussing with other reviewers, I am still concerned about the experiments. As a new field, it is okay not to compare other works, as they might be concurrent or not fully open sourced. However, consensus is reached that the work should adequately demonstrate that the proposed method is helpful for downstream applications while the current experiments part fail to do so. Thus, my final rating is Borderline reject.

Reviewer cvm42023-08-20

Thanks a lot for your clarifications and the efforts made to update the draft. While I no longer have any issues regarding the method/presentation, I believe the rebuttal does not fully address the concern regarding limited baselines in the analysis of the proposed model. The other reviewers have also commented on the insufficiency of the current experiments to demonstrate a high degree of realism and controllability with the proposed model, and I concur. Therefore, I am maintaining my rating of borderline reject.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC