Topological Game-Based Autonomous Driving: A Road Network Self-Play Decision Framework Inspired by the AlphaZero Paradigm

Engineering notes

Key topics: autonomous driving, motion planning, deployment, planning, control. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

This preprint proposes an AlphaZero-inspired topological game self-play decision framework for autonomous driving to address core defects of mainstream driving decision paradigms, including heavy reliance on massive real driving data, poor model interpretability, insufficient long-tail scenario coverage, and separated driving & parking algorithm architectures. The research decomposes road networks into 7 standardized road topological primitives plus 1 dedicated parking primitive to programmatically generate modular, resettable traffic simulation environments. It converts continuous vehicle motion control into discrete multi-objective combinatorial decision tasks on graph topologies. The framework leverages GNN and TopoMoE for cross-scale unified topological feature encoding of sparse road scenes and dense parking lots, adopts MCTS-UCT for multi-step forward-looking reasoning, and integrates Topology-GRPO to stabilize multi-agent self-play training. A hierarchical multi-objective reward system balances safety, traffic compliance, efficiency and ride comfort for both driving and parking tasks. High-quality trajectories are generated via simulation self-play without massive human-labeled data, and policy distillation enables lightweight deployment on vehicle-end devices. Different from industrial solutions that stack independent driving and parking modules, this work constructs an algorithm-native unified graph-based decision system for full-domain autonomous driving. It puts forward four core scientific hypotheses and staged verification schemes, delivers a five-stage experimental roadmap, and provides a low-data, low-computation alternative technical path for interpretable, generalizable intelligent driving motion planning.

Links and sources

Comments