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

Area of Science:

• Robotics
• Artificial Intelligence
• Autonomous Systems

Background:

• Unsignalized intersections present complex challenges for autonomous driving decision-making.
• Existing Monte Carlo Tree Search (MCTS) planners struggle with action coverage and lack integrated safety filters.

Purpose of the Study:

• To propose CP-LDS-MCTS, a novel decision-making framework for autonomous driving at unsignalized intersections.
• To improve safety, task completion, traffic efficiency, and control smoothness within computational limits.

Main Methods:

• CP-LDS-MCTS integrates Sobol low-discrepancy sampling for better action representation.
• It employs truncated Taylor control barrier function (TTCBF)-based safety pruning for pre-expansion action filtering.
• A policy-value composite scoring mechanism prioritizes safe, effective actions.

Main Results:

• CP-LDS-MCTS demonstrated superior performance in balancing safety, success rate, travel time, and smoothness in CARLA simulations.
• The method achieved a mean planning latency under 25 ms per step.
• It outperformed stronger baselines like PPO and MPC-CBF in complex traffic scenarios.

Conclusions:

• The proposed CP-LDS-MCTS framework offers a unified approach to continuous-action planning at unsignalized intersections.
• Jointly designing candidate coverage, safety screening, and value-aware expansion is crucial for real-time performance.
• The findings suggest a promising direction for enhancing autonomous driving safety and efficiency.