Beyond Self-Play: Hierarchical Reasoning for Continuous Motion in Closed-Loop Traffic Simulation

2026-05-09

AI Summary

Existing self-play reinforcement learning approaches struggle to model the socially aware behaviors of human drivers in closed-loop traffic simulation, often yielding policies that lack realism and safety. This work proposes a hierarchical framework in which a high-level Stackelberg multi-agent reinforcement learning module performs social interaction reasoning and generates intent-level commands, while a low-level continuous trajectory control module translates these commands into physically consistent driving actions. The approach innovatively integrates social awareness with continuous control and introduces a hybrid co-training strategy to mitigate distributional shift during closed-loop deployment. Evaluated on SUMO urban road networks, the method significantly improves driving smoothness and safety while maintaining traffic efficiency comparable to baseline approaches.

Abstract

Closed-loop traffic simulation requires agents that are both scalable and behaviorally realistic. Recent self-play reinforcement learning approaches demonstrate strong scalability, but their equilibrium strategies fail to capture the socially aware behaviors of real human drivers. We propose a hierarchical architecture that goes beyond self-play by combining high-level multi-agent interaction reasoning with low-level continuous trajectory realization. Specifically, a Stackelberg-style Multi-Agent Reinforcement Learning (MARL) module generates interaction-aware intention commands. These commands condition a low-level continuous motion module, translating the strategic intent into physically consistent, scene-responsive control sequences. To mitigate distribution shift in closed-loop deployment, we introduce a hybrid co-training scheme combining MARL with auxiliary recovery supervision. Experiments on a SUMO-based urban network demonstrate that the proposed framework achieves superior control smoothness and safety compared to self-play and passive imitation baselines, while maintaining competitive traffic efficiency.

Problem

• closed-loop traffic simulation
• self-play
• socially aware behavior
• behavioral realism
• multi-agent interaction

Innovation

• Hierarchical Reasoning
• Multi-Agent Reinforcement Learning
• Closed-Loop Simulation
• Continuous Motion Control
• Distribution Shift Mitigation