Principle:CARLA simulator Carla Scenario Execution
| Knowledge Sources | |
|---|---|
| Domains | Simulation, Recording |
| Last Updated | 2026-02-15 00:00 GMT |
Overview
Scenario execution is the process of running a simulation with diverse actors and traffic patterns to produce meaningful recorded data for analysis and replay.
Description
A recorded simulation is only as valuable as the scenario it captures. Scenario execution in the context of recording and replay refers to the deliberate orchestration of simulation conditions -- spawning vehicles and pedestrians, configuring traffic patterns, setting environmental conditions, and letting the simulation run -- to produce data that is useful for testing, debugging, or training purposes.
CARLA provides several mechanisms for scenario execution. The most common approach uses the generate_traffic.py example script, which spawns a configurable number of vehicles and walkers with autopilot behavior, creating a bustling traffic environment. More controlled scenarios can be crafted using the ScenarioRunner framework or through manual actor placement and control via the Python API.
The key consideration for recording scenarios is diversity and coverage. A good recorded scenario includes a variety of actor types (cars, trucks, motorcycles, bicycles, pedestrians), traffic situations (intersections, lane changes, overtaking, pedestrian crossings), and environmental conditions (weather, time of day). This ensures the recorded data is useful for a wide range of analysis tasks.
Usage
Execute simulation scenarios for recording when you need to:
- Generate traffic datasets containing diverse driving situations for algorithm testing.
- Reproduce specific scenarios such as near-miss situations or complex intersection interactions.
- Create benchmark recordings that capture a standardized set of traffic conditions for consistent evaluation.
- Populate recordings with enough activity to make replay analysis meaningful.
- Stress-test systems by creating high-density traffic scenarios that push sensor and planning systems to their limits.
Theoretical Basis
Scenario execution for recording is governed by the principle of representative data generation. The statistical properties of the recorded data should reflect the conditions under which the system under test is expected to operate. This includes the distribution of actor types, the frequency of different traffic events (merges, stops, turns), and the range of environmental conditions.
CARLA's traffic manager implements a distributed behavioral model where each autopilot vehicle makes independent decisions based on local perception of its surroundings. This emergent traffic behavior produces naturalistic interactions without requiring explicit scripting of every vehicle's trajectory. The traffic manager controls parameters such as target speed, following distance, lane-change probability, and right-of-way compliance, allowing the user to tune the aggressiveness and diversity of traffic behavior.
For walker (pedestrian) actors, CARLA uses a navigation mesh approach where walkers are assigned random destination points on the sidewalk network and navigate to them using the AI controller. The random seed for walker behavior can be set for reproducibility.
The combination of autonomous vehicle behavior and pedestrian navigation produces complex, multi-agent interactions that emerge organically during the simulation, yielding rich recorded data without requiring manual trajectory scripting.