A New Technical Framework for Safe and Efficient Left Turns in Autonomous Driving

Abstract

Unprotected left turns are still a difficult problem for autonomous driving because perception uncertainty, occlusion, and unpredictable road-user behaviour need to be handled in a short decision time. This paper proposes a risk-adaptive technical framework for integrating conflict-zone graph encoding, multimodal arrival-time prediction, dynamic risk allocation, chance-constrained gap selection and jerk-limited trajectory optimisation. The framework estimates the probability distribution of conflict-zone entry and clearance times, adjusts the admissible risk according to visibility, localisation uncertainty, road conditions and prediction calibration. 18,000 closed-loop simulation tests, 2,400 log-replay cases and 320 closed-course experiments were conducted for performance assessment. The proposed method had a 98.1% turn-completion rate and a 0.18% collision rate, and its collision rate was 0.27-0.71% for the four representative baselines. The median stop-line-to-clearance time was 7.6s, and the average intersection throughput reached 428 vehicles per hour per approach. The expected calibration error was 0.012, the median computational latency was 41ms, and 99.2% of the closed-course control cycles maintained a feasible backup trajectory. With combined sensing and behaviour disturbance, the collision-free survival rate after three seconds of continuous occlusion was 99.1%. As shown in the above results, adding probabilistic prediction and risk-constrained planning can improve the safety of left turns without adding much latency or other issues.