RADE: Learning Risk-Adjustable Driving Environment via Multi-Agent Conditional Diffusion

TL;DR

RADE tackles the challenge of evaluating autonomous-vehicle safety by generating statistically realistic, risk-adjustable multi-agent traffic scenes. It introduces a multi-agent diffusion framework conditioned on a PET-based risk level, and a tokenized dynamics check to guarantee physical plausibility, enabling scalable stress-testing without manually crafted adversaries. The method preserves realism across risk levels and shows that higher risk yields more safety-critical events and crashes in simulations, validated on the real-world rounD dataset. This approach offers a data-driven, scalable tool for robust AV safety evaluation and targeted scenario analysis in complex traffic environments.

Abstract

Generating safety-critical scenarios in high-fidelity simulations offers a promising and cost-effective approach for efficient testing of autonomous vehicles. Existing methods typically rely on manipulating a single vehicle's trajectory through sophisticated designed objectives to induce adversarial interactions, often at the cost of realism and scalability. In this work, we propose the Risk-Adjustable Driving Environment (RADE), a simulation framework that generates statistically realistic and risk-adjustable traffic scenes. Built upon a multi-agent diffusion architecture, RADE jointly models the behavior of all agents in the environment and conditions their trajectories on a surrogate risk measure. Unlike traditional adversarial methods, RADE learns risk-conditioned behaviors directly from data, preserving naturalistic multi-agent interactions with controllable risk levels. To ensure physical plausibility, we incorporate a tokenized dynamics check module that efficiently filters generated trajectories using a motion vocabulary. We validate RADE on the real-world rounD dataset, demonstrating that it preserves statistical realism across varying risk levels and naturally increases the likelihood of safety-critical events as the desired risk level grows up. Our results highlight RADE's potential as a scalable and realistic tool for AV safety evaluation.

Figures (8)

• Figure 1: The proposed RADE framework. Conditioned on the desired risk level, RADE takes joint current state of all the vehicles as the input, and generates raw trajectories using a multi-agent diffusion network, which conducts denoising steps over vehicle states. These trajectories are then refined by a tokenized dynamics check module, producing the final valid joint future trajectories.
• Figure 2: Network architecture based on the MADiff backbone zhu2024madiff. The model adopts a U-Net structure with multi-agent attention applied to skip-connected features, and receives a risk latent vector at each layer.
• Figure 3: Tokenized dynamics check. (a) Schematic for dynamics check. This module refines the raw trajectory by finding the closet motion token from a set of motion vocabulary. (b) Visualization of motion token vocabulary constructed from the rounD dataset krajewski2020round. The vocabulary contains $1024$ tokens, and here for better visualization, we show $512$ tokens.
• Figure 4: Snapshots of two scenarios under different risk levels. For each scenario, the left and right panels represent vehicle states at $t=0\,\mathrm{s}$ and $t=2\,\mathrm{s}$, respectively. In each case, RADE is used once to generate the future joint trajectories, visualized as colored lines. Vehicles exhibiting more risk-inducing behaviors are highlighted in red. The comparisons between low-risk (left columns) and high-risk (right columns) setups illustrate how RADE adjusts agent behavior based on the specified risk level.
• Figure 5: Statistical realism at different risk levels. Each metric is calculated by aggregating results from $10$ random-seeded experiments. The distance distribution is closely aligned with real-world data at different risk levels, and the speed distribution exhibits a similar overall trend.