How ut-adl/autoware_mini works
Main Components
• Localization Module – components under “nodes/localization” and the transformer helpers in “src/localization”. This part gathers sensor input (e.g., GNSS) and converts among coordinate frames.
• Global Planner – located in “nodes/planning/global”, which uses maps (e.g., Lanelet2) provided in “data/maps” to generate a high-level route.
• Local Planner – found in “nodes/planning/local” that, using the global path and sensor data, generates the actual local trajectory.
• Obstacle/Detection Modules – several nodes are distributed across “nodes/detection” for various sensor-based detection strategies (lidar_cluster, lidar_sfa, radar, and sensor fusion) as well as traffic light detection.
• Following/Control Module – components under “nodes/control” (both classic controllers like pure pursuit or Stanley follower) that convert planned trajectories into vehicle actuator commands.
• Platform Interfaces – under “nodes/platform” where files exist for simulation interfaces with Carla, real hardware (e.g., Lexus) and leaderboard systems.
• Helper and Utility Functions – in “src/helpers” and parts of “config”, which support common algorithms (geometry, transforms, timers, etc).
• External Integration – including ROS launch files (“launch” directory) for starting different configurations (planner simulation, bag replay, Carla simulation, and others) and support for external systems like Carla simulator, Scenario Runner, and vehicle platform.
Relationships and Interactions
• A clear flow from sensor input (simulated or recorded data from Carla, bag files, or actual sensor drivers) entering the perception/detection modules.
• That detection outputs are passed to the planning layer (both global and local planners) where the environment is interpreted and a route is produced.
Architectural Patterns
• Modular Design – emphasize that every capability (localization, planning, detection, and control) is implemented as independent ROS nodes.
• Separation of Concerns – clearly distinguish perception (different detection strategies), planning, and control.
• Extensibility – note that the use of configuration files (in “config”) and modular node design allows swapping different detection algorithms (e.g., lidar_cluster vs. lidar_sfa) and support for various simulation modes (simulation launch files, Carla integration, etc).
Relevant Technologies and Frameworks
• ROS (Noetic) is the backbone; include it as the central middleware facilitating node communication.
• Python plays a key role as the programming language for nodes.
• Autonomous driving components such as Lanelet2 maps and possibly ONNX Runtime for neural network-based detection (for GPU-enabled nodes).
• Integration with simulation tools like Carla and Scenario Runner.