CP-LDS-MCTS: A Decision-Making Method for Unsignalized Intersections Based on Low-Discrepancy Sampling and Safety Pruning

Unsignalized intersections pose a representative challenge for autonomous-driving decision-making because online planning must satisfy tightly coupled requirements for safety, task completion, traffic efficiency, and control smoothness under a limited computation budget. Existing continuous-action MCTS planners often suffer from sparse candidate-action coverage and from the absence of an internal safety filter before node expansion. To address these issues, this paper proposes CP-LDS-MCTS, a decision-making framework that coordinates Sobol low-discrepancy sampling, truncated Taylor control barrier function (TTCBF)-based safety pruning, and policy-value composite scoring within the expansion stage of Monte Carlo tree search. Sobol sampling improves candidate representativeness under a fixed sampling budget; TTCBF provides a local one-step screening rule that removes actions inconsistent with safety constraints before search resources are consumed; and composite scoring prioritizes safe actions that are simultaneously policy-consistent and value-promising. To clarify the methodological contribution, CP-LDS-MCTS is formulated as a unified expansion-stage design rather than a loose combination of independent modules. The revised manuscript further adds a local approximation-error discussion for the TTCBF truncation, a computational-complexity analysis, a real-time latency evaluation, statistical significance tests, and two stronger baselines, namely PPO and MPC-CBF. Experiments in CARLA Town03 under low-, medium-, and high-density traffic show that the proposed method achieves the best overall balance among safety, success rate, travel time, and control smoothness while maintaining a mean planning latency below 25 ms per step on the test platform. The resulting safety assurance is local rather than global, as TTCBF pruning performs a one-step approximation-based feasibility check within the expansion stage and is validated in simulation. These results suggest that candidate coverage, internal safety screening, and value-aware expansion should be designed jointly for real-time continuous-action planning at unsignalized intersections.