PaperlayerView exact revision on arXiv

Source-linked AI summary

Versatile Behavior Diffusion for Generalized Traffic Agent Simulation

Zhiyu Huang, Zixu Zhang, Ameya Vaidya, Yuxiao Chen, Chen Lv, Jaime Fernández Fisac

arXiv:2404.02524v3cs.RO

TL;DR

Traffic simulation can miss interactive behavior when replayed agents do not react, while marginal predictions may yield inconsistent joint scenes. VBD uses diffusion with behavior priors and user objectives to generate and guide multi-agent scenarios; it shows strong benchmark simulation performance, though a gap remains versus autoregressive models.

  • Problem

    Log replay leaves other traffic participants nonreactive, while marginal open-loop predictions can produce scene-inconsistent joint behavior; simulation also needs controllable scenarios for underrepresented dangerous situations.

    Source: §I · p. 1

  • Method

    VBD combines joint multi-agent diffusion with behavior-prior and user-objective guidance to generate or edit scenarios without retraining.

    Source: §I · p. 2, C. Controllable and Compositional… · p. 3, §VII · p. 8

  • Results

    VBD performs comparably to state-of-the-art generative models and outperforms motion-prediction models in simulation, with better realism metrics than SceneDiffuser.

    Source: §VI · p. 8

  • Takeaways & Limitations

    Guided generation supports customizable scenarios, including adversarial behaviors, while retaining joint multi-agent consistency and flexible inference-time control.

    Source: §VI · p. 8, §VII · p. 8

    A performance gap remains versus autoregressive models such as SMART and BehaviorGPT, and more denoising steps yield limited benefit under the realism metric.

    Source: §VI · p. 8

Abstract

from arXiv · show

Existing traffic simulation models often fall short in capturing the intricacies of real-world scenarios, particularly the interactive behaviors among multiple traffic participants, thereby limiting their utility in the evaluation and validation of autonomous driving systems. We introduce Versatile Behavior Diffusion (VBD), a novel traffic scenario generation framework based on diffusion generative models that synthesizes scene-consistent, realistic, and controllable multi-agent interactions. VBD achieves strong performance in closed-loop traffic simulation, generating scene-consistent agent behaviors that reflect complex agent interactions. A key capability of VBD is inference-time scenario editing through multi-step refinement, guided by behavior priors and model-based optimization objectives, enabling flexible and controllable behavior generation. Despite being trained on real-world traffic datasets with only normal conditions, we introduce conflict-prior and game-theoretic guidance approaches. These approaches enable the generation of interactive, customizable, or long-tail safety-critical scenarios, which are essential for comprehensive testing and validation of autonomous driving systems. Extensive experiments validate the effectiveness and versatility of VBD and highlight its promise as a foundational tool for advancing traffic simulation and autonomous vehicle development. Project website: https://sites.google.com/view/versatile-behavior-diffusion

I. INTRODUCTION

The introduction identifies scene inconsistency and limited controllability as gaps in traffic simulation, then presents VBD as a diffusion-based framework for realistic, controllable multi-agent behavior. Its related-work discussion distinguishes prior simulation, prediction, and diffusion approaches while motivating a more versatile framework.

I. INTRODUCTION

  • Log-replay testing leaves other participants nonreactive to the autonomous vehicle, limiting interactive scenarios and contributing to a simulation-to-reality gap.Examples include pedestrians entering crosswalks and vehicles negotiating right-of-way at four-way stops.Source: §I · p. 1

A. Traffic Simulation

  • Marginal open-loop predictions can produce scene-inconsistent joint rollouts, including unnatural interactions and collisions.Heuristic approaches, also struggle to scale to realistic traffic maneuvers.Source: §I · p. 1

I. INTRODUCTION

  • VBD is presented as a diffusion-based framework for realistic, controllable traffic-agent behavior, with strong closed-loop simulation performance and guidance schemes for user-specified scenarios.The contributions include compatibility with optimization objectives, behavior priors, and game-theoretic structures.Source: §I · p. 2

A. Traffic Simulation

  • As traffic simulation shifts toward learning-based methods to enhance realism and interactivity, examples include imitation learning, socially controllable behavior generation [24], and policies with parallel beam search to further enhance realism.Source: A. Traffic Simulation · p. 2

B. Behavior Prediction

  • Behavior prediction models estimate multimodal behaviors, but scalability and efficiency can limit their use in large scenarios; VBD uses behavior priors to control target-agent behavior –, [26]–[35].The cited work includes single-agent prediction and scene-level joint interaction prediction.Source: B. Behavior Prediction · p. 2

C. Diffusion Models for Traffic Simulation

III. PRELIMINARIES

The preliminaries formulate traffic scenario generation as conditional multi-agent trajectory optimization and connect its behavior distribution to maximum-entropy modeling. They then describe DDPM as a forward noising process followed by context-conditioned reverse denoising.

A. Problem Formulation

  • Traffic scenario generation is formulated as finite-horizon optimal control over joint agent trajectories and controls conditioned on maps, signals, and initial states.The dynamics define how controls produce trajectories, while the objective promotes realistic, feasible behavior.Source: A. Problem Formulation · p. 3
  • The conditional control-sequence distribution is represented as a Boltzmann distribution under Maximum Entropy IRL, and learning maximizes its conditional likelihood.The passage relates this representation to an energy-based model.Source: A. Problem Formulation · p. 3
  • Diffusion models address difficult score estimation by progressively noising data and learning the reverse process, which can act as an implicit planner toward low-cost, high-likelihood trajectories.The connection to trajectory optimization is framed as approximate stochastic gradient descent in score space.Source: A. Problem Formulation · p. 3

B. Denoising Diffusion Probabilistic Model (DDPM)

  • DDPM adds scheduled noise to clean control actions until they approach an isotropic Gaussian, then uses context-conditioned reverse diffusion to generate actions.The denoiser estimates clean actions from noisy samples, from which the preceding-step mean is derived.Source: B. Denoising Diffusion Probabilistic… · p. 3

C. Controllable and Compositional Generation

VBD supports user-specified scenario requirements without retraining by adding their objectives to the generation objective. During reverse diffusion, guidance modifies denoising steps using an objective gradient, with an alternative approximation evaluating the denoiser’s one-step result.

C. Controllable and Compositional Generation

  • User requirements such as agent goals, reference paths, speed-limit compliance, or collision avoidance are incorporated by adding an objective to the scenario-generation objective without retraining.The resulting conditional distribution composes the learned behavior distribution with the likelihood of the requirement.Source: C. Controllable and Compositional… · p. 3
  • During inference, predicted behavior priors or user-defined objectives guide denoising toward desired scenarios.Guidance alters the predicted mean at each reverse-diffusion step, with its strength controlled by λ.Source: Figure 2 · p. 4, C. Controllable and Compositional… · p. 4
  • Because noisy-action gradients can be inaccurate and unstable, an alternative evaluates the objective on the denoiser’s one-step generation result and differentiates through the denoiser.The gradient is taken with respect to the noisy actions.Source: C. Controllable and Compositional… · p. 4

IV. METHOD

VBD combines encoded scene context and individual-agent behavior priors with a denoiser that predicts joint multi-agent trajectories. Its scene representation includes agent histories, map polylines, and traffic lights, while the behavior predictor supplies multimodal trajectory options.

A. Model Structure

  • The denoiser predicts joint clean control sequences from scene context and noisy actions, using self-attention across agents and cross-attention to scene conditions.Its causal mask prevents information from future timesteps from being used at the current timestep during closed-loop rollout.Source: A. Model Structure · pp. 4–5
  • Scene context comprises agent histories, map polylines, and traffic lights, with agent, map, and signal features represented in local coordinates before encoding.Agent history includes position, heading, velocity, bounding-box dimensions, and agent type; map and signal inputs carry lane and light information.Source: A. Model Structure · pp. 4–5
  • The scene encoder combines GRU-encoded agent histories, MLP-encoded map and signal features, and query-centric Transformer layers to represent scene-element relationships.The resulting scene features have a uniform dimension.Source: A. Model Structure · p. 5
  • The behavior predictor decodes scene conditions into multimodal marginal trajectory distributions for individual agents, using endpoint anchors and assigning scores to candidate trajectories.Predicted action trajectories are converted to state trajectories through the dynamics function.Source: A. Model Structure · pp. 4–5

B. Model Training

VBD trains its encoder, denoiser, and predictor jointly, using noisy-trajectory denoising and an added behavior-prediction task. A log noise schedule and multi-task losses shape training.

B. Model Training

  • Joint training updates the encoder, denoiser, and predictor, while the denoising objective uses Smooth L1 loss between ground-truth and rolled-out trajectories.Source: B. Model Training · p. 5
  • The denoiser learns to recover clean trajectories from corrupted ones, with one sampled noise level applied jointly to all agents in a scene.Source: B. Model Training · p. 5
  • The log noise schedule keeps the signal-to-noise ratio sufficiently low to prevent inference degradation from shortcut learning and empirically outperforms the Cosine scheduler.Source: B. Model Training · p. 5
  • The behavior predictor stabilizes training, improves learning, and supplies individual-agent behavior priors; it selects modes by closest endpoint anchor when valid, otherwise by lowest average displacement error.Source: B. Model Training · p. 5
  • Prediction training uses Smooth L1 loss on selected trajectories and cross-entropy for mode scores; the loss is computed marginally, excludes invalid agents or time steps, and task losses are balanced by γ.Source: B. Model Training · pp. 5–6

V. EXPERIMENTAL SETUP

The setup maps agent actions and states through shared unicycle dynamics and evaluates VBD on Waymo data in closed-loop simulation. The authors report small trajectory reconstruction errors for the inverse dynamics model.

A. System Dynamics

  • The model uses shared unicycle dynamics to roll actions into states for vehicles, pedestrians, and cyclists, with agent states comprising position, yaw, and velocity.Source: A. System Dynamics · p. 6
  • Differentiable forward dynamics map predicted actions into states, while inverse dynamics derive approximate actions from ground-truth states.Source: A. System Dynamics · p. 6
  • Reconstructed trajectories have a displacement error of 0.221 m and a final displacement error of 0.511 m against ground-truth trajectories, indicating sufficiently accurate state-action reconstructions.Inverse dynamics provides an approximate control-space representation, while supervision is applied to state trajectories after forward rollout through differentiable forward dynamics.Source: A. System Dynamics · p. 6

B. Dataset

  • Experiments use WOMD’s 486,995 training and 44,097 validation scenarios, with 44,920 closed-loop test scenarios and 500 validation-interactive scenarios for additional experiments and ablations.WOMD and Waymax support longer-horizon, 8-second closed-loop simulation of every agent; the authors cite limitations of nuScenes and Argoverse for this evaluation setting.Source: B. Dataset · p. 6

C. Implementation Details

VBD combines joint multi-agent diffusion with inference-time guidance for closed-loop simulation and controllable scenario generation. Its Waymo Sim Agents performance is comparable to leading generative models, with better realism than SceneDiffuser.

C. Implementation Details

  • VBD encodes up to 64 agents, 256 map polylines, and 16 traffic lights, and generates future control sequences using a scene encoder, behavior predictor, and denoiser.The behavior predictor generates 64 candidate trajectories per agent; training uses 50 diffusion steps and 16 epochs.Source: C. Implementation Details · pp. 6–7

VI. CLOSED-LOOP TRAFFIC SIMULATION

  • Closed-loop evaluation rolls out eight-second trajectories with 1 Hz replanning; on the benchmark, VBD controls up to 64 agents while constant-velocity policies control others.Evaluation uses 32 independent rollouts per scene, over up to 128 agents, and drops past trajectories to address causal confusion.Source: §VI · p. 7
  • On the Waymo Sim Agents benchmark, VBD is comparable to state-of-the-art generative models, outperforms motion-prediction models, and has better realism than SceneDiffuser.It achieves this with fewer parameters than autoregressive generation models, although a performance gap remains against SMART and BehaviorGPT.Source: Table I · p. 8, §VI · p. 8

VII. CONTROLLABLE SCENARIO GENERATION

  • Guided sampling refines denoiser outputs with gradients of user-defined objectives, enabling inference-time scenario editing without retraining.The controllable-generation experiments use five gradient steps with scaling parameter λ = 0.1.Source: §VII · p. 8
  • Controllable-generation tests use 500 interactive validation scenes, 32 nearby agents, an eight-second horizon, 1 Hz replanning, and averages across three random seeds.Metrics include off-road incidents, collisions, wrong-way driving, kinematic infeasibility, and displacement-based log divergence.Source: §VII · p. 8

A. Composition with Behavior Priors

VBD uses goal priors to guide joint trajectory generation, seeking agent-specific behaviors while preserving scene consistency. Experiments report fewer collisions than direct marginal prediction and show guided changes to right-of-way behavior.

A. Composition with Behavior Priors

  • VBD selects target-agent goals from marginal behavior priors, then guides its denoiser to generate joint trajectories matching those goals.The other agents are not directly influenced by the guidance.Source: A. Composition with Behavior Priors · p. 9

Case Study 1: Scene-consistent Scenario Editing:

  • Guided diffusion reduces collision rates and better captures interactions than directly executing marginal predictions for target-agent goal-reaching behaviors.The approach can also alleviate off-road or wrong-way prior samples and produce scenarios that conform to scene context.Source: Case Study 1: Scene-consistent… · p. 9
  • Guiding Vehicle 5 to run a stop sign can produce a scenario in which Vehicle 1 yields, unlike the nominal rollout’s different scene-consistent ordering.Source: Case Study 1: Scene-consistent… · pp. 9–10

B. Composition with Cost Functions

Cost-based guidance steers VBD toward collision avoidance or a chosen driving style while preserving responsive multi-agent behavior. The framework also uses guidance to generate safety-critical scenarios without risk-specific training data.

Case Study 2: Enhancing Generation Quality:

  • Collision-avoidance cost guidance reduces collision rate and supports collision-free interactions in a challenging narrow-passage scenario.The paper reports the reduction in Table III and illustrates the narrow-passage interactions in Fig. 5.Source: Case Study 2: Enhancing Generation… · p. 10

Case Study 3: Modifying Driving Style:

VIII. GENERATION BEYOND DATA DISTRIBUTION

  • Conflict-prior and game-theoretic guidance generate realistic scenarios with collision risks for the AV without risk-specific driving data, retraining, or fine-tuning.These approaches are presented as ways to generate safety-critical scenarios beyond the training distribution.Source: §VIII · pp. 10–11

A. Safety-Critical Scenario Generation via Conflict Priors

Conflict-prior guidance uses predicted behavior modes to identify agents likely to conflict with the ego vehicle’s most likely plan, then guides diffusion with those behaviors. The reported scenarios are more realistic than direct trajectory rollout.

A. Safety-Critical Scenario Generation via Conflict Priors

  • The conflict-prior process identifies potentially unsafe agents from predicted behavior distributions and uses their selected priors to generate expected scenarios.The illustrated sequence predicts multimodal trajectories, selects a conflicting behavior, and generates a scenario with the diffusion policy.Source: A. Safety-Critical Scenario Generation… · p. 11, Figure 7 · p. 12
  • The method selects a high-probability behavior prior that conflicts with the ego vehicle’s most likely plan, then supplies it as guidance to the diffusion policy.Other agents remain under the policy without guidance, allowing them to respond to the guided situation.Source: A. Safety-Critical Scenario Generation… · p. 11, Figure 7 · p. 12
  • The reported generated behavior is more realistic and safety-critical than direct trajectory rollout.The result passage also mentions robustness against selected priors, but its scope is incomplete.Source: A. Safety-Critical Scenario Generation… · p. 12

B. Safety-Critical Generation via Game-theoretic Guidance

The paper combines VBD with game-theoretic guidance for interactive adversarial scenarios and studies design choices that affect simulation behavior. These include rollout-based training, history dropout, and a co-trained behavior predictor.

B. Safety-Critical Generation via Game-theoretic Guidance

  • Game-theoretic-guided diffusion proactively optimizes both pursuer and evader actions, producing interactive adversarial maneuvers in merge and lane-change scenes.Unlike approaches that sample against fixed pursued-vehicle trajectories, this method optimizes both agents’ actions.Source: B. Safety-Critical Generation via… · p. 12

A. Training Objective for Behavior Generation

  • Trajectory-rollout supervision produces coherent, realistic traffic scenarios, whereas open-loop supervision on controls or noise fails to capture temporal dynamics.The comparison used the same architecture for 20 epochs; the paper attributes the need for rollouts to errors accumulating across sequential decisions.Source: A. Training Objective for Behavior… · p. 12

B. Influence of History Encoding

  • Training with 50% history dropout and supplying only the current joint state at inference outperforms full-history encoding across all reported metrics.The full-history baseline can over-accelerate or decelerate in closed-loop simulation and shows high displacement errors.Source: B. Influence of History Encoding · p. 12, Table IV · p. 13

C. Integration of Behavior Prediction

D. Noise Scheduling Design

The paper reports that a logarithmic noise schedule supports diverse scenarios without compromising quality, while DDIM can accelerate inference with a quality-efficiency tradeoff. Its conclusion highlights broad task coverage and identifies runtime efficiency and domain extension as future work.

D. Noise Scheduling Design

  • Logarithmic noise scheduling maintains a suitable signal-to-noise ratio through diffusion and yields diverse scenarios without compromising generation quality.The paper attributes the schedule to the need for scene-context-dependent action denoising.Source: D. Noise Scheduling Design · pp. 13–14
  • Additional denoising steps expand feasible outcomes and support behavior control, while average ADE and FDE rise and minADE and minFDE decline.The paper reports satisfactory quality after the first step on collision rate, off-road rate, ADE, and FDE, with later steps affecting the fidelity-diversity balance.Source: D. Noise Scheduling Design · p. 14

E. Accelerating Inference

  • Five or ten DDIM steps offer a favorable balance between generation quality and real-time performance, with a feasible 1 Hz replan frequency.Runtime grows linearly with the number of diffusion steps; guided sampling can reduce collisions and improve realism without compromising computational efficiency.Source: E. Accelerating Inference · pp. 14–15, Table V · p. 14
  • With the same architecture and training steps, rollout-trajectory supervision improves planning quality over training on noise or control sequences.The figure describes the noise-trained model as producing meaningless behaviors and the control-trained model as misaligned with map features.Source: Figure 10 · p. 15

X. CONCLUSIONS

  • The conclusion reports effectiveness across traffic simulation tasks, including safety-critical scenarios, and identifies runtime efficiency and broader traffic domains as future-work areas.Other proposed directions include integration into AV planning and testing pipelines and context-aware generation using large language models.Source: §X · p. 15
Loading 2404.02524v3…