Note
In compliance with IROS double-blind review policy, the core EAMP planning algorithm code will be released in full upon potential acceptance of the paper. The current repository provides the complete logistics emergency simulation benchmark and a baseline RDA controller so that reviewers and researchers can reproduce the experimental scenarios described in the paper.
EAMP introduces a high-fidelity simulation benchmark for evaluating autonomous planning and control of last-mile logistics delivery vehicles operating in dynamic urban emergency scenarios. Built on top of the CARLA simulator and ROS, the benchmark features:
- Realistic logistics delivery scenarios on CARLA's Town10HD map with a compact delivery vehicle (Microlino) navigating through mixed traffic including pedestrians, motorcycles, and bicycles.
- Three difficulty levels (Easy / Medium / Hard) with progressively increasing numbers of dynamic obstacles, adverse weather conditions, and complex interactions.
- A* global path planning with a local lane-graph overtake planner that generates smooth, collision-free reference trajectories.
- MPC-RDA baseline controller — a Model Predictive Control framework integrated with Reciprocal Distance Avoidance (RDA) for safe, real-time obstacle avoidance and trajectory tracking.
- Rich sensor suite including 64-channel LiDAR, RGB camera, and semantic segmentation camera attached to the ego vehicle.
- Configurable weather system supporting 13 weather presets (clear, cloudy, rain, fog, etc.) for robustness evaluation.
The ego delivery vehicle autonomously navigates through urban obstacles using MPC-RDA control. Left: RGB camera and semantic segmentation views. Right: 64-channel LiDAR point cloud in RViz.
Each scenario spawns an ego delivery vehicle at a fixed start point and defines a goal location. Dynamic actors are triggered based on proximity to the ego vehicle, creating realistic emergency interactions:
| Difficulty | Pedestrians | Motorcycles / Bicycles | Weather | Description |
|---|---|---|---|---|
| Easy | 1 | 1 | Cloudy | Single pedestrian crossing + one motorcycle with lane change |
| Medium | 2 | 2 | Cloudy | Two pedestrians crossing + motorcycle and bicycle interactions |
| Hard | 3 | 3 | Cloudy | Dense mixed traffic with multiple concurrent dynamic obstacles |
All obstacle actors feature distance-triggered activation — they remain stationary until the ego vehicle approaches within a configurable radius, then execute pre-programmed motion patterns (e.g., pedestrian crossing, motorcycle lane change).
┌─────────────────────────────────────────────────────────┐
│ ROS Launch File │
├──────────┬──────────┬───────────┬──────────┬────────────┤
│ Scenario │ Global │ MPC │ Obstacle │ Spectator │
│ Spawner │ Planner │Controller │ Actors │ Camera │
├──────────┴──────────┴───────────┴──────────┴────────────┤
│ CARLA-ROS Bridge (carla_ros_bridge) │
├─────────────────────────────────────────────────────────┤
│ CARLA Simulator (Town10HD_Opt) │
└─────────────────────────────────────────────────────────┘
| Module | Description |
|---|---|
scenario_spawner_*.py |
Spawns ego vehicle, obstacles, pedestrians, and attaches sensors (LiDAR, cameras) |
global_planner.py |
A*-based global path planning + local lane-graph overtake planner with quintic smoothing |
controller.py |
MPC-RDA controller that reads ego state from CARLA, detects obstacles, and computes control commands |
walker.py / motor.py |
Behavior controllers for pedestrians and motorcycles with distance-triggered activation |
rda_planner/ |
Core MPC solver with RDA-based obstacle avoidance using CVXPY and ADMM optimization |
spectator_controller.py |
Third-person camera following the ego vehicle |
vehicle_data_recorder.py |
Records vehicle telemetry (speed, yaw, throttle, steering, sideslip) for analysis |
| Requirement | Version |
|---|---|
| Ubuntu | 20.04 LTS |
| CUDA | 12.2 |
| Python | 3.8 |
| ROS | Noetic |
| CARLA | 0.9.13 |
| CARLA-ROS-Bridge | Compatible with CARLA 0.9.13 |
numpy
cvxpy
pathos
PyYAML
carla==0.9.13
Follow the official CARLA installation guide.
mkdir -p ~/workspace
cd ~/workspace
git clone https://anonymous.4open.science/r/EAMP-307B/
cd ~/workspace/EAMP/src
git clone https://github.com/carla-simulator/ros-bridge.git
cd ~/workspace/EAMP
# Install all ROS system dependencies (ros_numpy, cv_bridge, pcl_ros, etc.)
rosdep install --from-paths src --ignore-src -r -y
# Install Python dependencies
pip install -r requirements.txt
catkin_makeThe
requirements.txtwill automatically install all required Python dependencies including the CARLA Python API (carla==0.9.13).
Ifpip install carla==0.9.13fails, you can also install it manually from the.egg/.whlfile bundled with your CARLA installation:pip install /path/to/CARLA_0.9.13/PythonAPI/carla/dist/carla-0.9.13-cpXX-cpXX-linux_x86_64.whl
cd $CARLA_ROOT
./CarlaUE4.sh
$CARLA_ROOTis the path to your CARLA 0.9.13 installation directory.
source ~/workspace/EAMP/devel/setup.bash
roslaunch carla_ros_bridge carla_ros_bridge.launch town:=Town10HD_Opt synchronous_mode:=False timeout:=20000source ~/workspace/EAMP/devel/setup.bash
# Easy mode — minimal obstacles
roslaunch logistics easy.launch
# Medium mode — moderate obstacle density
roslaunch logistics medium.launch
# Hard mode — dense mixed traffic
roslaunch logistics hard.launchThe ego vehicle will automatically plan a global path, navigate through dynamic obstacles using the MPC-RDA controller, and stop upon reaching the goal.
EAMP/
├── data/
│ ├── dataset/ # VLM evaluation dataset (test.jsonl)
│ ├── image/ # Front-camera RGB images from CARLA
│ └── README.md
├── src/
│ ├── logistics/ # Main ROS package
│ │ ├── launch/ # Launch files for Easy/Medium/Hard scenarios
│ │ ├── scripts/ # ROS nodes (spawner, planner, controller, actors)
│ │ ├── src/rda_planner/ # MPC-RDA solver (CVXPY + ADMM)
│ │ ├── msg/ # Custom ROS messages (Obstacle, PerceivedObject)
│ │ ├── srv/ # Custom ROS services (AnalyzeScene, RequestOvertake)
│ │ └── rviz/ # RViz visualization config
│ └── ros-bridge/ # CARLA-ROS bridge (cloned separately)
├── requirements.txt
└── README.md
| Component | Status |
|---|---|
| Logistics emergency benchmark (scenarios, obstacle actors, sensor setup) | Released |
| MPC-RDA baseline controller | Released |
| A* + local lane-graph overtake planner | Released |
This work builds upon the following open-source projects: