Autonomous Vehicle Decision-Making Systems

The autonomous driving decision-making system serves as the "brain" of an autonomous vehicle. It receives environmental data from the perception system, determines the immediate driving action required, plans a dedicated route, and relays commands to the vehicle control system for execution.

🧠 Core Workflow

The decision-making system typically operates across several progressive layers:

1. Behavioural Decision (Behaviour Decision): Addresses the macro question of "what should be done right now"—such as following the lead vehicle, executing an overtaking manoeuvre, or giving way. This stage dictates the vehicle's overarching driving strategy.

2. Motion Planning: Once the behavioural strategy is set, this stage calculates a specific, safe, and smooth driving trajectory from the current position to the target waypoint. The trajectory accounts for road curvature, the relative position and speed of surrounding traffic, and passenger ride comfort.

3. Closed-Loop Control (Control): Translates the planned trajectory into precise, real-time steering angle, throttle, and braking inputs, continuously making adjustments based on real-time vehicle dynamics to keep the vehicle tracking accurately along the intended path.

💡 Implementation Approaches: Three Mainstream Technical Architectures

Executing this decision-making framework primarily relies on three technical routes, each with distinct strengths and trade-offs:

1. Modular Systems (Modular System): A traditional and widely adopted architecture that decouples behavioural decisions, motion planning, and control into standalone modules interconnected via deterministic rules or models. It offers strong interpretability and straightforward fault diagnosis during calibration; however, cascading errors across modules can accumulate, degrading overall system performance.

2. End-to-End Systems (End-to-End System): This approach utilises a deep neural network to map raw sensor feeds (such as camera imagery) directly into driving actuation commands (such as steering angle inputs). Its principal strength lies in eliminating cascading errors to enable global optimisation, offering a higher theoretical performance ceiling. However, its "black box" nature results in poor interpretability, making decision rationales difficult to audit—a critical liability in safety-stringent automotive applications.

3. Optimisation-Based Systems (Optimization-based System): Serving as a middle ground between the two approaches, this framework defines complex mathematical "cost functions" to quantify safety, efficiency, and ride comfort, solving for the optimal path and driving strategy. Offering verifiable safety and predictable vehicle behaviour, it is widely considered the most viable solution for large-scale series production. Key technical implementations include Model Predictive Control (MPC) and Artificial Potential Field (APF) methods.

🚀 Frontier Technological Evolution

Decision-making technology is advancing rapidly, driven by key technical trends:

• Imitation Learning and Reinforcement Learning: Training models on extensive human driving datasets allows AI to adopt experienced driving behaviours (imitation learning). Alternatively, reinforcement learning enables the system to interact with the environment via trial and error, deriving optimal policies through reward and penalty feedback; research from Tongji University underscores the vast potential of deep reinforcement learning in automated planning and decision-making.

• Deep Integration of Prediction and Planning: Conventional architectures sequentially predicted surrounding traffic trajectories before planning vehicle movement. Current trends integrate both into a unified pipeline—such as Task-Unified Planning Systems—aligning predictive models directly with end driving objectives to significantly enhance obstacle avoidance and complex manoeuvre completion rates.

• Uncertainty-Aware Decision-Making: Real-world driving environments present high unpredictability, such as sudden cut-ins by adjacent vehicles. Advanced frameworks like the Hierarchical Partially Observable Markov Decision Process (POMDP) adopted by Hi-Drive explicitly account for environmental uncertainties during computation to deliver more robust, fail-safe path planning.

In summary, autonomous vehicle decision-making systems are shifting from rule-based, highly interpretable modular designs towards a data-driven, intelligent unified era that harmonises functional safety with optimal performance.