TLDR: CARJAN is a novel open-source tool that integrates the AJAN multi-agent framework with the CARLA driving simulator to provide a user-friendly, visual interface for generating and simulating complex urban traffic scenarios. It enables transparent modeling of intelligent agent behaviors using SPARQL Behavior Trees and offers live simulation with real-time behavior tree visualization, simplifying the development and testing of autonomous systems and traffic interactions.
Creating realistic and interactive virtual traffic scenarios for testing autonomous vehicles, understanding pedestrian behavior, and simulating urban environments has long been a complex challenge. Researchers and developers often struggle with user-friendly tools that can model diverse interacting agents like pedestrians, cyclists, and self-driving cars in a clear and explainable way.
Addressing this need, a new tool called CARJAN has been introduced. Developed by Leonard Frank Neis, André Antakli, and Matthias Klusch from the German Research Center for Artificial Intelligence (DFKI), CARJAN offers a novel approach to semi-automated generation and simulation of traffic scenarios. It integrates the multi-agent engineering framework AJAN with the popular driving simulator CARLA, providing a unified platform for scenario design and execution.
What is CARJAN?
CARJAN stands out by offering a visual user interface that simplifies the modeling, storage, and maintenance of traffic scenario layouts. Unlike some existing frameworks that require extensive coding, CARJAN leverages SPARQL Behavior Trees for defining the decision-making and interactions of agents in dynamic simulations within CARLA. This declarative approach makes agent behaviors more transparent and easier to understand.
The Foundation: AJAN and CARLA
CARJAN is built upon two key technologies. The first is AJAN (Accessible Java Agent Nucleus), a robust framework for engineering multi-agent systems. AJAN enables agents to perform semantic reasoning over data and execute behaviors defined using SPARQL-extended Behavior Trees. It includes a knowledge graph management system and a runtime environment, with the AJAN-Editor providing a web-based, graphical interface for modeling and debugging agent behaviors through drag-and-drop.
The second pillar is CARLA, an open-source simulation platform widely used for autonomous driving research. CARLA provides realistic urban environments and a Python API for creating scenarios and controlling agents. While CARLA is powerful, its native scenario modeling can be static, code-intensive, and prone to errors. CARJAN aims to overcome these limitations by offering a more structured and visual approach.
How CARJAN Works
The architecture of CARJAN seamlessly connects AJAN and CARLA through a middleware component called `carjanService`, which uses the Flask web framework. Users visually model scenarios using a grid-based layout in CARJAN’s GUI, placing dynamic and static entities. Intelligent agents, such as pedestrians or vehicles, are automatically assigned to and initialized within AJAN, where their behaviors are defined using SPARQL Behavior Trees.
For scenario generation, CARJAN provides an intuitive GUI where users can select from map templates (like T-junctions or intersections) and position entities via drag-and-drop. Paths for agents are defined using waypoints, and “decision boxes” can be set up as interactive zones that trigger agent reactions upon collision. Agent behaviors are modeled with a predefined set of actions and semantic knowledge, allowing for dynamic decision-making. For instance, a pedestrian agent can decide when to cross a road based on the distance and behavior of an approaching car, as observed in the simulation.
A significant advantage of CARJAN is its automated translation process. Once a scenario is modeled, converting it into a CARLA-compatible format for simulation is a one-click operation. The `carjanService` handles all necessary conversions and coordination, making iterative experimentation much faster and less tedious than traditional scripting methods.
Integrated Simulation and Explainability
CARJAN offers an integrated simulation environment where users can watch their modeled scenarios come to life in CARLA, directly within the CARJAN GUI. A unique feature is the `LiveBehavior` tool, which provides real-time visualization of an agent’s behavior tree execution. This allows developers to see the current status of each node in the tree (inactive, executed, successful, or failed) through color-coding, making the agent’s decision-making process highly explainable. A runtime console further displays execution logs and agent status, helping users correlate observed behavior with the underlying logic.
This integrated approach not only simplifies the creation and testing of complex traffic scenarios but also enhances the transparency of agent behaviors, which is crucial for developing and validating autonomous systems. For more technical details, you can refer to the full research paper here.
Also Read:
- Smart Supply Chains: How AI Makes Optimization Clear and Interactive
- Unpacking Epistemic Forgetting in AI: A Unified Framework for Knowledge Management
Looking Ahead
CARJAN represents a significant step forward in the field of traffic scenario generation and simulation. As an open-source tool, it empowers developers to transparently model, simulate, and test various learning and planning methods for diverse agent behaviors within a single, user-friendly graphical interface. Future versions are expected to include improved support for benchmark maintenance and more scene layout and agent templates.
