Principle:CARLA simulator Carla Navigation Agent Architecture

Knowledge Sources	CARLA Simulator CARLA Documentation Russell & Norvig, "Artificial Intelligence: A Modern Approach", Chapter 2 - Intelligent Agents, 4th Ed., 2020 Paden et al., "A Survey of Motion Planning and Control Techniques Adopted in Self-Driving Vehicles", IEEE T-IV, 2016
Domains	Autonomous Driving, Agent Architecture, Motion Planning
Last Updated	2026-02-15 00:00 GMT

Overview

The Navigation Agent Architecture in CARLA defines a hierarchical software design pattern for autonomous vehicle control. A base agent class (BasicAgent) provides core capabilities -- global route planning, local waypoint tracking, obstacle detection, and PID-based vehicle control -- while a specialized subclass (BehaviorAgent) extends this foundation with personality-driven driving behaviors such as cautious, normal, and aggressive profiles.

Description

The architecture follows a layered decomposition common in robotics and autonomous driving systems:

1. Global Planning Layer: Uses a GlobalRoutePlanner to compute a route from origin to destination over the road network graph. The route is expressed as a sequence of (carla.Waypoint, RoadOption) tuples.
2. Local Planning Layer: A LocalPlanner maintains a waypoint queue (deque) and selects the next target waypoint for the vehicle to track. It converts the global plan into immediate control targets.
3. Perception Layer: The agent queries the CARLA world for nearby vehicles, pedestrians, and traffic lights. Geometric corridor-based obstacle detection determines whether hazards exist along the planned route.
4. Control Layer: A VehiclePIDController with separate longitudinal (throttle/brake) and lateral (steering) PID controllers generates the final carla.VehicleControl command.

BasicAgent composes these layers and exposes a simple run_step() loop that:

• Detects obstacles (vehicles, walkers, traffic lights)
• Issues emergency stop if hazards are present
• Otherwise delegates to the local planner for PID-based waypoint tracking

BehaviorAgent inherits from BasicAgent and adds:

• Behavior profiles (Cautious, Normal, Aggressive) with tunable parameters for safety time, speed limits, and following distances
• Tailgating logic for slow lead vehicles
• Car-following manager with Time-To-Collision (TTC) based speed adaptation
• Lane-change decisions at intersections

Usage

Use the navigation agent architecture when you need to:

• Rapidly prototype an autonomous driving pipeline in CARLA without writing control logic from scratch
• Compare different driving behaviors (cautious vs. aggressive) under identical scenarios
• Build a baseline agent for benchmarking more sophisticated planning algorithms (e.g., reinforcement learning or model-predictive control)
• Teach or demonstrate the sense-plan-act paradigm in a realistic driving simulator

Theoretical Basis

The CARLA agent architecture implements the classical sense-plan-act paradigm from robotics:

loop:
    percepts = sense(world)          # Query vehicles, walkers, traffic lights
    if hazard_detected(percepts):
        action = emergency_stop()
    else:
        plan = local_planner.next_waypoint()
        action = pid_controller(plan)
    vehicle.apply_control(action)

Hierarchical Decomposition

Layer	Component	Responsibility
Global Planning	GlobalRoutePlanner	Compute route over road network graph (A* search)
Local Planning	LocalPlanner	Maintain waypoint queue, select next target waypoint
Perception	BasicAgent._vehicle_obstacle_detected, _affected_by_traffic_light	Detect obstacles and traffic signals along route corridor
Control	VehiclePIDController	Convert waypoint error into throttle, brake, and steering commands

Inheritance Hierarchy

BasicAgent
    |-- GlobalRoutePlanner   (has-a)
    |-- LocalPlanner         (has-a)
    |-- VehiclePIDController (via LocalPlanner)
    |
    +-- BehaviorAgent (is-a)
            |-- Cautious / Normal / Aggressive behavior profiles
            |-- car_following_manager()
            |-- tailgating logic

The base class provides safe, functional defaults (obey traffic lights, stop for obstacles, follow lane at target speed). The subclass overrides run_step() to inject behavior-specific decision-making while reusing the inherited perception and control infrastructure.

Related Pages

Implemented By

• Implementation:CARLA_simulator_Carla_BasicAgent_Init