Dual-Branch Trajectory Decision-Making for Autonomous Driving in Intersection Left-Turn Scenarios

Intersection left-turn decision-making remains challenging for autonomous vehicles because the ego vehicle must select safe and efficient maneuvers under uncertain surrounding-vehicle motion, constrained conflict-region space, and strict requirements for online computation. To address this problem, this study proposes a dual-branch trajectory decision-making method driven by interaction-aware spatiotemporal prediction. A level-k GameFormer predictor generates multimodal ego candidates and surrounding-vehicle predictions. These candidates are then converted into a compact PPO action space for efficient primary decision-making. Potential conflict vehicles are filtered by their spatial relationship with the ego reference path. A spatial safety-margin criterion activates bounded MCTS correction only when the selected PPO trajectory is predicted to be unsafe. Closed-loop experiments on 500 nuPlan intersection left-turn scenarios show that the proposed method achieves an 87.9% success rate and a 6.7% collision rate. The average episode-level minimum distance dmin is 5.78 m, and the average step time is 29.3 ms. Compared with the GameFormer Top-1 baseline, the proposed method increases the success rate by 15.7 percentage points and reduces the collision rate by 67.3%. Compared with GameFormer + PPO + TTC-MCTS, it achieves a lower collision rate and a shorter average step time under the paired 20-seed evaluation protocol. These results indicate that prediction-guided candidate selection, explicit spatial safety evaluation, and selective local correction can improve left-turn decision safety while maintaining real-time feasibility under the tested closed-loop protocol.