Transformer-Based Predictive Motion Planning at Signalized Intersections: A Symmetry-Breaking Perspective in a SUMO–CARLA Co-Simulation Environment

Autonomous vehicles operating at signalized intersections face fundamental challenges arising from queue dynamics, signal-phase transitions, and tightly coupled multi-vehicle interactions. Conventional motion-planning methods, which rely primarily on instantaneous perception, are inherently reactive and struggle to reason about short-term traffic evolution. This paper presents a Transformer-based predictive motion-planning framework that embeds short-term traffic state prediction directly into the structure of the planning problem. A lightweight spatial–temporal Transformer model is designed to forecast traffic occupancy, queue evolution, and interaction patterns using historical trajectories, signal-phase information, and road topology. By converting predicted traffic dynamics into explicit spatial–temporal constraints, a hierarchical motion planner jointly optimizes path geometry and speed profiles through dynamically constructed feasible corridors. The proposed framework is evaluated using a joint SUMO–CARLA simulation platform under realistic traffic conditions derived from real-world datasets, including pNEUMA and CitySim. The experimental results across straight-through, queueing, and turning scenarios show that prediction-aware planning significantly reduces high-risk driving time and intersection travel time while maintaining stable real-time computational performance. Beyond scenario-level improvements, the results indicate that transforming traffic prediction into planning constraints provides a generalizable paradigm for proactive, feasibility-aware autonomous driving at signalized intersections. From a methodological perspective, the proposed framework can be interpreted through the lens of symmetry and asymmetry in intelligent transportation systems: the conventional symmetric decoupling between prediction and planning modules is deliberately broken by embedding predicted traffic states as time-varying, directionally asymmetric constraints, while the permutation symmetry of the multi-head attention mechanism is preserved over lane-segment tokens to provide a structured inductive bias for traffic state forecasting. This symmetry-aware design highlights how controlled symmetry breaking in modeling and optimization can yield safer, more efficient, and more adaptive autonomous driving behaviors in signalized urban environments.