Trajectory Planning

Trajectory planning is a core component within autonomous driving decision-making systems. Once the behavioural decision-making layer determines "what to do" (such as car-following or overtaking), trajectory planning addresses "how to execute it"—formulating a driving path that satisfies safety, ride comfort, and the vehicle's physical handling limits.

In simple terms: behavioural decision-making sets the direction, while trajectory planning plots the route. The outputted trajectory defines not only the destination path, but also the precise speed, timing, and vehicle attitude required at every point along the way.

Core Challenge: From "Static Path" to "Dynamic Trajectory"

The complexity of trajectory planning stems from its fundamental difference compared to conventional static "path planning":

• Path Planning: Focuses strictly on the geometric route without factoring in time—such as plotting a collision-free line from point A to point B on a navigation map.

• Trajectory Planning: Integrates the dimension of time into the path, calculating the vehicle's exact position, velocity, and acceleration at every single moment.

Incorporating time makes this an exceptionally complex task (classified mathematically as an NP-hard problem). The system must coordinate both spatial navigation and speed modulation simultaneously to adapt to dynamic traffic conditions. For instance, when evading a vehicle that suddenly cuts in, the system has to adjust steering angle and vehicle speed in tandem to produce a trajectory that is both safe and smooth.

Key Technical Approaches

Currently, research into trajectory planning algorithms centres around three main technical pathways, each presenting distinct strengths and trade-offs:

1. Conventional Planning Algorithms

• Representative Methods: Graph search (e.g., A*, Dijkstra), sampling-based approaches (e.g., RRT), artificial potential field methods, and parametric curve generation.

• Advantages: Highly explainable, with easily verifiable safety and feasibility, serving as the backbone for many series-production systems.

• Disadvantages: Computational efficiency can become a bottleneck in highly complex dynamic environments, and generated trajectories often require secondary smoothing.

2. Learning-Based Approaches

• Representative Methods: Reinforcement learning, imitation learning, etc.

• Advantages: Strong capability in tackling complex scenarios that are challenging to model via conventional algorithms, capable of human-like driving strategies with a higher performance ceiling.

• Disadvantages: Limited explainability due to the "black box" nature; safety cannot be mathematically guaranteed; requires massive volumes of high-quality training data.

3. Optimisation and Model Predictive Control (MPC)

• Representative Methods: Formulating trajectory planning as an optimal control problem resolved through frameworks such as MPC.

• Advantages: Simultaneously handles path and speed profiling while accounting for real-time vehicle dynamics constraints, making it ideal for synchronous decision-making and trajectory tracking in complex scenarios.

• Disadvantages: Computationally intensive, demanding high-end onboard computing power and stringent algorithm real-time responsiveness.

Future Trends

The current industry trend clearly leans towards fusing the safety and determinism of conventional algorithms with the intelligence of learning-based models into hybrid architectures. For example, reinforcement learning manages complex behavioural decisions, while conventional optimisation algorithms generate the final, fail-safe trajectory. Furthermore, multimodal foundation models such as Vision-Language-Action (VLA) architectures are being explored for trajectory planning, empowering autonomous systems with deeper contextual understanding and reasoning capabilities.