Preferential Terrain Navigation

Overview

Preferential Terrain Navigation is a research project focused on developing an autonomous navigation system for robots capable of traversing semi-structured, off-road environments. Unlike many existing navigation systems that prioritize the shortest path, this system seeks to emulate human-like decision-making by considering terrain preferences. For example, a human would typically choose a sidewalk over a snowy path, even if the sidewalk is longer.

The system integrates terrain recognition with optimal control to achieve this goal. The approach consists of two key steps:

1. Terrain Recognition and Mapping: Using computer vision to analyze the environment and generate a terrain map with cost values that incorporates human-like terrain preferences and obstacles.
2. Path Planning and Refinement: Using the terrain map to plan a global path through trajectory optimization and refining the route using model predictive control (MPC).

This repository contains the implementation of the system using the CARLA simulator and the Robot Operating System (ROS).

---

Features

• Terrain Recognition: Leveraging computer vision and segmentation techniques to classify terrain types and create a preference-based terrain map. Key steps consist of segmenting the camera images mounted on the robot, applying inverse perspective mapping to get the BEV image, and then convert that into a grid map with cost values associated with obstacles and terrain preferences.
• Path Planning: Utilizing path planning algorithms and trajectory optimization to compute global paths that align with human-like terrain preferences.
• Path Refinement: Applying model predictive control (MPC) to adjust the planned route dynamically in response to real-time conditions.
• Obstacle Avoidance: Combine MPC with Control Barrier Function (CBF) to provide guarantee of obstacle avoidance.
• Simulation Environment: Using CARLA to simulate unstructured environments and test navigation algorithms.

---

Prerequisites

CARLA Simulator:

Download CARLA version 0.9.13.

Robot Operating System (ROS):

Install ROS (tested with ROS Noetic).

Eigen:

• Installation:

sudo apt install libeigen3-dev

Python:

Python 3.8 or higher.

Python Packages:

• Installation:

pip install -r requirements.txt

Grid_Map Library:

• Installation:

sudo apt install ros-<your-ros-distro>-grid-map*

Additional dependencies are managed through the included CMakeLists.txt and ROS packages.

---

Getting Started

Useful Commands

Start CARLA Simulator

cd ~/CARLA_0.9.13
./CarlaUE4.sh -prefernvidia

Start ROS Bridge

roslaunch carla_ros_bridge carla_ros_bridge.launch
roslaunch carla_spawn_objects carla_spawn_objects.launch
roslaunch carla_manual_control carla_manual_control.launch
# or equivalently
roslaunch carla_ros_bridge carla_ros_bridge_with_example_ego_vehicle.launch

Start Ackermann Message Converter

roslaunch carla_ackermann_control carla_ackermann_control.launch

Start Waypoint Publisher and AD Demo

roslaunch carla_waypoint_publisher carla_waypoint_publisher.launch

Start AD Demo with Scenario

roslaunch carla_ad_demo carla_ad_demo_with_scenario.launch

Start RViz

<node pkg="rviz" type="rviz" name="rviz" args="-d $(find carla_ad_demo)/config/carla_ad_demo.rviz" required="true" output="screen">
    <remap from="carla/ego_vehicle/spectator_pose" to="/carla/ego_vehicle/rgb_view/control/set_transform"/>
</node>

Start Segmentation Node

rosrun segmentation segmentation_node.py

Start Bird’s Eye View (BEV) Node

rosrun bev ipm_node.py --gt

Start Bird’s Eye View (BEV) Node

rosrun bev ipm_node.py --gt

---

Useful Topics

Sensor Topics

• IMU: /carla/ego_vehicle/imu (Message Type: sensor_msgs/Imu)
• GNSS: /carla/ego_vehicle/gnss (Message Type: sensor_msgs/NavSatFix)
• Odometry: /carla/ego_vehicle/odometry (Message Type: nav_msgs/Odometry)
• Front RGB Camera: /carla/ego_vehicle/camera/rgb/front/image_color (Message Type: sensor_msgs/Image)
• Speedometer: /carla/ego_vehicle/speedometer (Message Type: std_msgs/Float64)
• Front Radar: /carla/ego_vehicle/radar_front (Message Type: carla_msgs/CarlaRadar)
• LiDAR: /carla/ego_vehicle/lidar (Message Type: sensor_msgs/PointCloud2)
• Laser Scan: /scan (Message Type: sensor_msgs/LaserScan) – added using the pointcloud_to_laserscan package.

---

Repository Structure

PreferentialTerrainNavigation
├── src
│   ├── bev                  # Bird’s Eye View (BEV) generation
│   ├── segmentation         # Terrain segmentation
│   ├── mapping              # Grid Map Creation from Segmented BEV Image
│   ├── planning             # Path planning and MPC modules
│   ├── ros-bridge           # Integration with ROS and CARLA
│   ├── CMakeLists.txt       # Build configuration

Work in Progress

• Smooth local planned path and integrate into controller.
• Implement elevation mapping and integrate into terrain map cost.
• Train segmentation model on combined datasets (cityscape, rellis, carla).
• Implement control barrier function (CBF) and integrate with mpc.
• Drone assistance: make drone follow robot from above with a mounted camera looking down.

---

How to Contribute

1. Fork the repository and create a new branch for your feature or bug fix.
2. Ensure your code adheres to the existing style and structure.
3. Submit a pull request with a detailed description of your changes.

---

Acknowledgments

• CARLA Simulator
• ROS
• The research community exploring autonomous navigation in unstructured environments.

---

Contact

For questions or collaboration, please contact Simon Li at xi.yang.li@mail.mcgill.ca.