Path Planning

Path planning is a pivotal module in autonomous driving systems responsible for determining the exact route taken from Point A to Point B. Operating downstream of "behavioural decision-making" (which determines actions such as following or overtaking), it translates abstract driving intentions into a concrete, feasible, and optimal geometric trajectory.

Role and Objectives: The "Navigator" of Autonomous Driving

If autonomous driving is likened to a road trip, the decision-making layer acts as the team leader determining "where to go", the control layer functions as the driver handling the throttle and steering, while path planning serves as the navigator charting the course from the map.

Its primary objective is to compute the optimal driving route from origin to destination based on road network data from high-definition (HD) maps and the Global Positioning System (GPS), factored in with real-time traffic conditions and road regulations. This "optimum" can represent the shortest distance, quickest travel time, lowest energy consumption, or highest safety margin.

Core Methodologies: From "Graph Search" to "Intelligent Learning"

Mathematically, path planning can be conceptualised as an optimal path search problem across a "road network graph" (where roads are mapped as nodes and edges). The primary approaches to solving this include:

• Classical Search Algorithms: This remains the most established and mature methodology. Exemplified by the Dijkstra algorithm and the A* algorithm, these techniques systematically traverse nodes within the road network graph to identify the path with the lowest cost function (such as the shortest distance). Offering robust interpretability and deterministic results, they form the bedrock of many production-ready solutions; however, computational efficiency can become a bottleneck in highly complex or dynamically evolving environments.

• Heuristic and Bio-inspired Algorithms: To enhance search efficiency and accommodate intricate constraints, researchers have adopted heuristic methods such as the Ant Colony Optimisation (ACO) algorithm. By emulating biological behaviours in nature through parallel multi-path exploration and pheromone accumulation, these algorithms can converge on superior routes more efficiently, excelling particularly in path smoothing and multi-objective optimisation tasks.

• Deep Reinforcement Learning: Representing a cutting-edge frontier in recent years, this approach trains AI models through extensive trial-and-error simulations, enabling the system to autonomously formulate optimal path decisions within complex, dynamic traffic flows. While holding immense potential to resolve intricate scenarios that defy traditional mathematical modelling, it demands substantial training datasets and compute power, alongside persistent challenges in interpretability.

Trajectory Flow Management: The Evolution from "Navigation" to "End-to-End Service"

A distinct industry trend is the transition of path planning from basic route calculation towards highly refined trajectory flow management. This means the system must tackle not only broad routing, but also granular, complex operational tasks—such as selecting the correct ETC lane through toll plazas, autonomously locating vacant parking bays in unfamiliar car parks without prior map learning, and executing voice prompts such as "find the exit".

This necessitates a deep integration of path planning with semantic understanding, environmental perception, and chassis control modules, driving the evolution from rudimentary point-to-point guidance into an end-to-end, seamless automated service across complex operational domains like toll plazas, car parks, and commercial precincts.