Gaussian Process-based Traversability Analysis for Terrain Mapless Navigation

• Related Paper: arXiv Preprint
• Video Demo: YouTube

This work has been accepted for publication at 2024 IEEE International Conference on Robotics and Automation. (ICRA 2024)

Table of Contents

• Installation
• Simulation Setup
• Running
• Module Parameters
• Results
• Citation

Installation

The code was tested on Ubuntu 20.04, using ROS Noetic, and Python 3.8.10.

1. Clone GitHub Repositories:

cd ~/catkin_ws/src
git clone https://github.com/abeleinin/gp-navigation.git 
git clone https://github.com/merose/diff_drive.git

Clone the current repository and merose/diff_drive, which provides a Python differential-drive controller API.

2. Install Python and ROS dependencies:

cd gp-navigation
pip install -r requirements.txt
sudo apt install ros-noetic-tf2-sensor-msgs

3. Download the model files for simulation environments a and b from this OneDrive link into the gp-navigation package directory.

4. Build workspace:

cd ../..
catkin_make
source devel/setup.bash

Simulation Setup

Environments

We tested our code in 2 Gazebo environments shown above. Originally, environment A is Map 1 from the our baselines repo (PUTN). Environment B was chosen as a more challenging terrain with less flat areas.

Robot Platform

The Clearpath Robotics Husky robot was used during testing. We also include launch files to run our method with the Jackal robot. The LiDAR sensor we use is the Velodyne VLP-16.

Groundtruth localization in Gazebo

To use the ground truth localization for Husky in Gazebo, add the p3d_base_controller plugin in the Husky xacro file husky_description/urdf/husky.urdf.xacro

<gazebo>
  <plugin name="p3d_base_controller" filename="libgazebo_ros_p3d.so">
    <alwaysOn>true</alwaysOn>
    <updateRate>60.0</updateRate>
    <bodyName>base_link</bodyName>
    <topicName>ground_truth/state</topicName>
    <gaussianNoise>0.01</gaussianNoise>
    <frameName>world</frameName>
    <xyzOffsets>0 0 0</xyzOffsets>
    <rpyOffsets>0 0 0</rpyOffsets>
  </plugin>
</gazebo>

• Note: you need to add the p3d_base_controller plugin into jackal.urdf.xacro file if you are using Jackal instead of Husky.

A-LOAM localization

We also tested our method with aloam localization to make a fair comparison with our baseline method (PUTN). Since, the repo was created prior to ROS Noetic's release, but we we're still able to make it work with the following dependency version. We found that ceres solve version 2.0.0 works and using the following installation guide.

Running

Here is a demo of how you'd launch and run Task 3 from our paper:

1. Spawn Husky into Environment B at the starting location:

roslaunch gp_navigation husky_gazebo.launch env:=env_b x:=-7 y:=-7 z:=-5.2

Use the env argument to chose the environment (default: env_a).

2. Next, use the corresponding gp_navigation_{robot}_{env}.launch file to launch our method with the appropriate robot and environment configurations:

roslaunch gp_navigation gp_navigation_husky_env_b.launch

In RViz you can use the 2D Nav Goal tool to select a goal. To send more precise goals you can publish a Pose to the topic /plan_path_to_goal as shown below:

rostopic pub --once /plan_path_to_goal geometry_msgs/PoseStamped "header:
  seq: 0
  stamp:
    secs: 0
    nsecs: 0
  frame_id: 'world'
pose:
  position:
    x: 10.0
    y: 18.0
    z: 0.0
  orientation:
    x: 0.0
    y: 0.0
    z: 0.0
    w: 1.0"

Module Parameters

All the parameters for our method is stored in the config/ directory. Click to expand for parameter descriptions:

GP Mapping Module Parameters

GP Mapping configs located in config/gp_module/.

Name	Description	Default
length_in_x	map size in the x-axis	10.0
length_in_y	map size in the y-axis	10.0
resolution	Pixel density of maps grids	0.2
inducing_points	Subset of points to train GP model on	500

Path Planning Module Parameters

Path Planning configs located in config/path_planning/.

Name	Description	Default
step_len	Length of RRT* branching	0.5
iter_max	Maximum interations of RRT*	1000
radius	Circle about the robot, usually half the map size	5.0
trav_limit	Threshold for traversability safety	Determines how agressive to plan.
replan_dist	Distance to sub-goal before initiating replanning	Map size and compute dependent.
critical	Limitation on what the robot can safely traverse	Environment and robot dependent.
trav_weights	Weighting coefficients to determine their impact on traversability calculation	Environment and robot dependent. Weights should sum to 1.
sub_goal_weights	Weighting for sub-goal selection method	Environment and robot dependent. Weights should sum to 1.

Results

Our baseline comparison was PUTN: A Plane-fitting based Uneven Terrain Navigation Framework. The authors released their code here on GitHub: jianzhuozhuTHU/putn. We we're able to out perform their method in Task 2, which involved traversing the more challenging Environment B.

Citation

If you find this code or our work valuable, please cite:

@article{leininger2024,
  title={Gaussian Process-based Traversability Analysis for Terrain Mapless Navigation},
  author={Leininger, Abe and Ali, Mahmoud and Jardali, Hassan and Liu, Lantao},
  journal={arXiv preprint arXiv:2403.19010},
  year={2024}
}