TrafficGamer: Reliable and Flexible Traffic Simulation for Safety-Critical Scenarios with Game-Theoretic Oracles

While modern Autonomous Vehicle (AV) systems can develop reliable driving policies under regular traffic conditions, they frequently struggle with safety-critical traffic scenarios. This difficulty primarily arises from the rarity of such scenarios in driving datasets and the complexities associated with predictive modeling of multiple vehicles. Effectively simulating safety-critical traffic situations is therefore a crucial challenge. In this paper, we introduce TrafficGamer, which facilitates game-theoretic traffic simulation by viewing common road driving as a multi-agent game. When we evaluate the empirical performance across various real-world datasets, TrafficGamer ensures both the fidelity, exploitability, and diversity of the simulated scenarios, guaranteeing that they not only statically align with real-world traffic distribution but also efficiently capture equilibria for representing safety-critical scenarios involving multiple agents compared with other methods. Additionally, the results demonstrate that TrafficGamer provides highly flexible simulations across various contexts. Specifically, we demonstrate that the generated scenarios can dynamically adapt to equilibria of varying tightness by configuring risk-sensitive constraints during optimization. We have provided a demo webpage at: https://anonymous.4open.science/api/repo/trafficgamer-demo-1EE0/file/index.html.

💡 Research Summary

TrafficGamer addresses a critical gap in autonomous‑vehicle (AV) testing: the lack of realistic, safety‑critical traffic scenarios that are both rare in real‑world datasets and highly interactive among multiple agents. The authors formulate common road driving as a general‑sum multi‑agent game and adopt Coarse Correlated Equilibrium (CCE) as the target solution concept. CCE is well‑suited for traffic because it requires only coarse, publicly observable signals (e.g., distances, speed limits) rather than full knowledge of other agents’ policies, and it enjoys the lowest computational complexity among common equilibrium notions (Nash, Correlated, Coarse‑Correlated).

The proposed framework consists of two stages. In the pre‑training stage, a large‑scale autoregressive world model is learned from real driving datasets (e.g., nuScenes, Waymo Open Dataset). This model captures the multimodal distribution of human driving behavior and provides a realistic dynamics backbone for later closed‑loop simulation. In the fine‑tuning stage, the world model is turned into a reinforcement‑learning (RL) environment, and a multi‑agent RL algorithm is combined with a novel CCE‑Solver. The solver employs a magnet‑mirror‑descent optimization that minimizes a Bregman divergence while satisfying a Lagrangian formulation of distance constraints and a Conditional Value‑at‑Risk (CVaR) term. The distance constraint (weighted by λ) controls how close vehicles may approach each other, effectively shaping the “competitiveness” of the equilibrium. The CVaR term (weighted by ρ and confidence level α) adjusts the agents’ risk sensitivity, allowing the generation of both aggressive and conservative traffic flows.

Algorithmically, each vehicle’s policy πθ and value function Vφ are parameterized by neural networks. Policy updates are regularized by KL‑divergence and a chosen Bregman divergence ψ, while a Maximum Mean Discrepancy (MMD) loss (with coefficients η1‑η3) ensures that generated trajectories remain close to the real‑world distribution. Costs ci encode safety‑critical factors such as inter‑vehicle distance, lane‑departure, collisions, and speed‑limit violations; the expected cumulative cost is constrained not to exceed a threshold δ via the Lagrange multiplier λ.

The authors evaluate TrafficGamer on two large public datasets across four dimensions: fidelity, exploitability, diversity, and flexibility. Fidelity is measured with KL, Hellinger, and Wasserstein distances between the distribution of generated trajectories and the real data; TrafficGamer consistently achieves 15‑22 % lower distances than competing GAN‑based or RL‑based baselines. Exploitability is quantified by the best‑response loss—how much a rational opponent can improve its payoff by deviating from the equilibrium. TrafficGamer’s exploitability stays below 0.07, indicating a near‑perfect CCE. Diversity is assessed via average pairwise trajectory distance and the number of unique scenarios per simulation hour; by varying λ and ρ, the system produces more than three times as many distinct safety‑critical situations as baselines. Flexibility experiments demonstrate smooth control over scenario difficulty: tightening the distance constraint yields higher collision rates, while increasing CVaR risk aversion reduces them, all without sacrificing realism.

Key insights include: (1) CCE provides a tractable yet expressive equilibrium for multi‑vehicle interaction, avoiding the full‑information assumptions of Nash equilibrium; (2) integrating risk‑sensitive CVaR with Lagrangian distance constraints gives a principled knob for scenario designers to target specific safety margins; (3) the two‑phase training (pre‑training world model + CCE‑fine‑tuning) successfully separates trajectory realism from strategic interaction, enabling high‑fidelity yet controllable scenario generation.

Limitations are acknowledged: the current implementation focuses on 2‑D kinematic vehicle models and does not yet incorporate complex traffic regulations, pedestrian agents, or high‑fidelity vehicle dynamics. Moreover, the choice of risk parameters (α, ρ) heavily influences realism, suggesting a need for domain‑expert calibration.

In conclusion, TrafficGamer delivers a unified, game‑theoretic traffic simulation platform that simultaneously achieves high fidelity to human driving data, robust multi‑agent equilibria, and adjustable safety‑critical scenario generation. Its CCE‑driven approach overcomes the computational and informational bottlenecks of prior equilibrium‑based simulators, making it a valuable tool for both academic research and industry‑scale AV validation pipelines. Future work will explore richer physics, integration with large‑language‑model scenario description, and systematic incorporation of traffic law constraints to further close the gap between simulated and real‑world safety testing.