PCT Planner

Overview

This is an implementation of paper Efficient Global Navigational Planning in 3-D Structures Based on Point Cloud Tomography (accepted by TMECH). It provides a highly efficient and extensible global navigation framework based on a tomographic understanding of the environment to navigate ground robots in multi-layer structures.

Demonstrations: pct_planner

Citing

If you use PCT Planner, please cite the following paper:

Efficient Global Navigational Planning in 3-D Structures Based on Point Cloud Tomography

@ARTICLE{yang2024efficient,
  author={Yang, Bowen and Cheng, Jie and Xue, Bohuan and Jiao, Jianhao and Liu, Ming},
  journal={IEEE/ASME Transactions on Mechatronics}, 
  title={Efficient Global Navigational Planning in 3-D Structures Based on Point Cloud Tomography}, 
  year={2024},
  volume={},
  number={},
  pages={1-12}
}

Prerequisites

Environment

• Ubuntu >= 20.04
• ROS2 Humble with desktop-full installation (ros-humble-desktop)
• CUDA >= 11.7

Python

• Python >= 3.10 (bundled with Ubuntu 22.04)
• CuPy matching your CUDA version
• Open3d

Setup

1. Source ROS2 Humble

Add this to your ~/.bashrc (run once):

echo "source /opt/ros/humble/setup.bash" >> ~/.bashrc
source ~/.bashrc

2. Create Virtual Environment

ROS2 Humble requires Python 3.10. The venv must use --system-site-packages so that ROS2 Python packages (rclpy, sensor_msgs_py, etc.) remain accessible:

python3.10 -m venv pct_env --system-site-packages
source pct_env/bin/activate

To activate the environment in future sessions:

source pct_env/bin/activate

3. Install Python Dependencies

pip install open3d "numpy<2"
# Install CuPy matching your CUDA version (check with: nvidia-smi)
# CUDA 12.x:
# pip install cupy-cuda12x
# CUDA 13.x:
pip install cupy-cuda13x

The planner native modules currently require NumPy 1.x. If you installed NumPy 2.x, downgrade first:

pip install "numpy==1.26.4"

Build & Install

Inside the package, there are two modules: the point cloud tomography module for tomogram reconstruction (in tomography/) and the planner module for path planning and optimization (in planner/). You only need to build the planner module before use. In planner/, run build_thirdparty.sh first and then run build.sh.

cd planner/
./build_thirdparty.sh
./build.sh

Run Examples

Three example scenarios are provided: "Spiral", "Building", and "Plaza".

• "Spiral": A spiral overpass scenario released in the 3D2M planner.
• "Building": A multi-layer indoor scenario with various stairs, slopes, overhangs and obstacles.
• "Plaza": A complex outdoor plaza for repeated trajectory generation evaluation.

Tomogram Construction

To plan in a scenario, first you need to construct the scene tomogram using the pcd file.

• Unzip the pcd files in rsc/pcd/pcd_files.zip to rsc/pcd/.
• For scene "Spiral", you can download the pcd file from 3D2M planner spiral0.3_2.pcd.
• Start RViz2 with the provided config:

rviz2 -d rsc/rviz/pct_ros.rviz

• Activate your virtual environment and, in tomography/scripts/, run tomography.py with the --scene argument:

source pct_env/bin/activate
cd tomography/scripts/
python3 tomography.py --scene Spiral
python3 tomography.py --scene Isaacsim

• The generated tomogram is visualized as ROS2 PointCloud2 message in RViz2 and saved in rsc/tomogram/.

Trajectory Generation

After the tomogram is constructed, you can run the trajectory generation example.

• In planner/scripts/, run plan.py with the --scene argument:

source pct_env/bin/activate
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/YOUR/DIRECTORY/TO/PCT_planner/planner/lib/3rdparty/gtsam-4.1.1/install/lib:/YOUR/DIRECTORY/TO/PCT_planner/planner/lib/build/src/common/smoothing
cd planner/scripts/
python3 plan.py --scene Spiral


export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/wendy/Documents/GitHub/PCT_planner/planner/lib/3rdparty/gtsam-4.1.1/install/lib:/home/wendy/Documents/GitHub/PCT_planner/planner/lib/build/src/common/smoothing
cd planner/scripts/
python3 plan.py --scene Isaacsim

• The generated trajectory is visualized as ROS2 Path message in RViz2.

Live Global Planning (robot pose + RViz2 goal)

plan_global.py subscribes to the robot's live localization (via TF map → base_link) as the start position and to RViz2's "2D Goal Pose" tool as the goal position, then publishes the planned path to /pct_path.

source pct_env/bin/activate
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/wendy/Documents/GitHub/PCT_planner/planner/lib/3rdparty/gtsam-4.1.1/install/lib:/home/wendy/Documents/GitHub/PCT_planner/planner/lib/build/src/common/smoothing
cd planner/scripts/

# Goal layer by index (0 = ground floor)
python3 plan_global.py --scene Isaacsim --goal_layer 0

# Goal layer by height in meters (auto-converted to layer)
python3 plan_global.py --scene Isaacsim --goal_z 1.5

Arguments

Argument	Default	Description
--scene	Isaacsim	Scene name (selects tomogram file)
--tomo_file	(from scene)	Override tomogram filename (without .pickle)
--goal_layer	0	Goal layer index (0 = ground floor)
--goal_z	(not set)	Goal height in meters — auto-converted to layer, overrides --goal_layer
--start_layer	0	Fallback start layer if TF Z is unavailable
--start_z	(from TF)	Override start height in meters instead of using TF Z
--robot_frame	base_link	TF frame of the robot base
--map_frame	map	TF map frame
--goal_topic	/goal_pose	RViz2 "2D Goal Pose" topic
--pose_topic	/localization	Fallback pose topic (nav_msgs/Odometry in map frame) used when TF lookup fails

Notes

• --goal_layer uses a layer index (integer). To find how many layers your tomogram has, check the terminal output of tomography.py for Num slices simp: N.
• --goal_z uses meters in the map frame. The node converts it to the nearest layer using slice_h0 and slice_dh from the tomogram.
• A new path is replanned every time you set a new "2D Goal Pose" in RViz2.

License

The source code is released under GPLv2 license.

For commercial use, please contact Bowen Yang byangar@connect.ust.hk.

Quick Start (ROS2 Humble)

Open three terminals. In each, first run:

source ~/.bashrc
source /path/to/PCT_planner/pct_env/bin/activate

Terminal 1 — Tomogram Construction

cd tomography/scripts/
python3 tomography.py --scene Spiral

Terminal 2 — RViz2 Visualization

cd /path/to/PCT_planner
rviz2 -d rsc/rviz/pct_ros.rviz

Terminal 3 — Trajectory Planning

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/path/to/PCT_planner/planner/lib/3rdparty/gtsam-4.1.1/install/lib:/path/to/PCT_planner/planner/lib/build/src/common/smoothing
cd planner/scripts/
python3 plan.py --scene Spiral

python3 plan.py --scene Isaacsim

Configuration

Planner Input (planner/scripts/plan.py)

Parameter	Description
start_pos	Start position in meters [x, y] (map frame)
end_pos	Goal position in meters [x, y] (map frame)
start_layer	Start layer index (0 = ground floor)
end_layer	Goal layer index

The Z coordinate (height) is automatically determined by the tomogram layer. You only specify the 2D position and which layer to start/end on.

Tomogram Slicing (tomography/config/scene_*.py)

Parameter	Description	Default
map.resolution	Grid cell size in meters	0.20
map.slice_dh	Height per initial slice (meters)	0.5
map.ground_h	Ground plane height	0.0

The number of final layers is automatically determined by layer simplification: initial slices that don't contain unique traversable surfaces are merged. Smaller slice_dh → more initial slices → potentially more final layers.

Traversability Parameters (tomography/config/scene_*.py)

Parameter	Description
trav.slope_max	Maximum traversable slope (radians)
trav.step_max	Maximum step height (meters)
trav.safe_margin	Obstacle clearance distance (meters)
trav.inflation	Cost inflation radius (meters)

Troubleshooting

Issue	Fix
cudaErrorUnknown / cuInit() = 999	GPU driver in bad state. Run sudo reboot
CycloneDDS enp130s0 does not match	Set CYCLONEDDS_URI in ~/.bashrc to use loopback: export CYCLONEDDS_URI='<CycloneDDS><Domain><General><Interfaces><NetworkInterface name="lo"/></Interfaces></General></Domain></CycloneDDS>'
RViz2 shows empty topics	Start RViz2 before or at the same time as the publisher nodes (QoS: TRANSIENT_LOCAL)
Segmentation fault in planner	Ensure libcommon_smoothing.so path is in LD_LIBRARY_PATH
source ~/.bashrc deactivates venv	Re-activate with source pct_env/bin/activate after sourcing bashrc
Segmentation fault or RuntimeError around pickle / init_map	Use numpy==1.26.4, rebuild planner/, and regenerate the tomogram pickle with tomography.py. Old NumPy 2 generated pickles are not safe to load in the planner.