Elevation Mapping cupy

Overview

This is a ROS package for elevation mapping on GPU. The elevation mapping code is written in python and uses cupy for GPU computation. The plane segmentation is done independently and runs on CPU. When the plane segmentation is generated, local convex approximations of the terrain can be efficiently generated.

Citing

Takahiro Miki, Lorenz Wellhausen, Ruben Grandia, Fabian Jenelten, Timon Homberger, Marco Hutter
Elevation Mapping for Locomotion and Navigation using GPU arXiv

@misc{mikielevation2022,
  doi = {10.48550/ARXIV.2204.12876},
  author = {Miki, Takahiro and Wellhausen, Lorenz and Grandia, Ruben and Jenelten, Fabian and Homberger, Timon and Hutter, Marco},
  keywords = {Robotics (cs.RO), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Elevation Mapping for Locomotion and Navigation using GPU},
  publisher = {International Conference on Intelligent Robots and Systems (IROS)},
  year = {2022},
}

Installation

CUDA & cuDNN

The tested versions are CUDA10.2, 11.6

CUDA
cuDNN.

Check how to install here.

Python dependencies

You will need

• cupy
• numpy
• scipy
• shapely==1.7.1

For traversability filter, either of

• torch
• chainer

Optionally, OpenCV for inpainting filter.

• opencv-python

Install numpy, scipy, shapely, opencv-python with the following command.

pip3 install -r requirements.txt

Cupy

cupy can be installed with specific CUDA versions. (On jetson, only "from source" i.e. pip install cupy could work)

For CUDA 10.2 pip install cupy-cuda102

For CUDA 11.0 pip install cupy-cuda110

For CUDA 11.1 pip install cupy-cuda111

For CUDA 11.2 pip install cupy-cuda112

For CUDA 11.3 pip install cupy-cuda113

For CUDA 11.4 pip install cupy-cuda114

For CUDA 11.5 pip install cupy-cuda115

For CUDA 11.6 pip install cupy-cuda116

(Install CuPy from source) % pip install cupy

Traversability filter

You can choose either pytorch, or chainer to run the CNN based traversability filter.
Install by following the official documents.

• torch
• chainer

Pytorch uses ~2GB more GPU memory than Chainer, but runs a bit faster.
Use parameter use_chainer to select which backend to use.

ROS package dependencies

• pybind11_catkin
• grid_map

sudo apt install ros-noetic-pybind11-catkin
sudo apt install ros-noetic-grid-map-core ros-noetic-grid-map-msgs

On Jetson

CUDA CuDNN

CUDA and cuDNN can be installed via apt. It comes with nvidia-jetpack. The tested version is jetpack 4.5 with L4T 32.5.0.

python dependencies

On jetson, you need the version for its CPU arch:

wget https://nvidia.box.com/shared/static/p57jwntv436lfrd78inwl7iml6p13fzh.whl -O torch-1.8.0-cp36-cp36m-linux_aarch64.whl
pip3 install Cython
pip3 install numpy==1.19.5 torch-1.8.0-cp36-cp36m-linux_aarch64.whl

Also, you need to install cupy with

pip3 install cupy

This builds the packages from source so it would take time.

ROS dependencies

• pybind11_catkin
• grid_map

sudo apt install ros-melodic-pybind11-catkin
sudo apt install ros-melodic-grid-map-core ros-melodic-grid-map-msgs

Also, on jetson you need fortran (should already be installed).

sudo apt install gfortran

If the Jetson is set up with Jetpack 4.5 with ROS Melodic the following package is additionally required:

git clone git@github.com:ros/filters.git -b noetic-devel

Plane segmentation dependencies

OpenCV

sudo apt install libopencv-dev

Eigen

sudo apt install libeigen3-dev

CGAL

CGAL5 is required. It will be automatically downloaded and installed into the catkin workspace by the cgal5_catkin package. Make sure you have the third-party libaries installed on you machine:

sudo apt install libgmp-dev
sudo apt install libmpfr-dev
sudo apt install libboost-all-dev

Usage

Build

catkin build elevation_mapping_cupy
catkin build convex_plane_decomposition_ros

Errors

If you get error such as

Make Error at /usr/share/cmake-3.16/Modules/FindPackageHandleStandardArgs.cmake:146 (message):
  Could NOT find PythonInterp: Found unsuitable version "2.7.18", but
  required is at least "3" (found /usr/bin/python)

Build with option.

catkin build elevation_mapping_cupy -DPYTHON_EXECUTABLE=$(which python3)

Run

Basic usage.

roslaunch elevation_mapping_cupy elevation_mapping_cupy.launch

For the plane segmentation node

roslaunch convex_plane_decomposition_ros convex_plane_decomposition.launch

Errors

If you build with the install flag under ros melodic, you might get issues with the modules not found:

terminate called after throwing an instance of 'pybind11::error_already_set'
  what():  ModuleNotFoundError: No module named 'elevation_mapping_cupy'

This is because python3 modules are installed into a different location.

This can be fixed by including also the python3 modules location in the PYTHONPATH by adding following line into the launch file:

<env name="PYTHONPATH" value="<path_to_your_install>/lib/python3/dist-packages:$(env PYTHONPATH)" />

Run TurtleBot example

First, install turtlebot simulation.

sudo apt install ros-noetic-turtlebot3-gazebo ros-noetic-turtlebot3-teleop

Then, you can run the examples. For the basic version:

export TURTLEBOT3_MODEL=waffle
roslaunch elevation_mapping_cupy turtlesim_example.launch

Or, for the version including plane segmentation:

catkin build convex_plane_decomposition_ros
export TURTLEBOT3_MODEL=waffle
roslaunch elevation_mapping_cupy turtlesim_segmentation_example.launch

To control the robot with a keyboard, a new terminal window needs to be opened.
Then run

export TURTLEBOT3_MODEL=waffle
roslaunch turtlebot3_teleop turtlebot3_teleop_key.launch

Velocity inputs can be sent to the robot by pressing the keys a, w, d, x. To stop the robot completely, press s.

Subscribed Topics

• topics specified in pointcloud_topics in elevation_mapping_cupy/config/parameters.yaml ([sensor_msgs/PointCloud2])

  The distance measurements.

• /tf ([tf/tfMessage])

  The transformation tree.

• The plane segmentation node subscribes to an elevation map topic ([grid_map_msg/GridMap]). This can be configured in convex_plane_decomposition_ros/config/parameters.yaml

Published Topics

For elevation_mapping_cupy, topics are published as set in the rosparam.
You can specify which layers to publish in which fps.

Under publishers, you can specify the topic_name, layers basic_layers and fps.

publishers:
  your_topic_name:
    layers: [ 'list_of_layer_names', 'layer1', 'layer2' ]             # Choose from 'elevation', 'variance', 'traversability', 'time' + plugin layers
    basic_layers: [ 'list of basic layers', 'layer1' ]                # basic_layers for valid cell computation (e.g. Rviz): Choose a subset of `layers`.
    fps: 5.0                                                        # Publish rate. Use smaller value than `map_acquire_fps`.

Example setting in config/parameters.yaml.

• elevation_map_raw ([grid_map_msg/GridMap])

  The entire elevation map.

• elevation_map_recordable ([grid_map_msg/GridMap])

  The entire elevation map with slower update rate for visualization and logging.

• elevation_map_filter ([grid_map_msg/GridMap])

  The filtered maps using plugins.

The plane segmentation node publishes the following:

• planar_terrain ([convex_plane_decomposition_msgs/PlanarTerrain])

  A custom message that contains the full segmentation as a map together with the boundary information.

• filtered_map ([grid_map_msg/GridMap])

  A grid map message to visualize the segmentation and some intermediate results. This information is also part of planar_terrain.

• boundaries ([visualization_msgs/MarkerArray])

  A set of polygons that trace the boundaries of the segmented region. Holes and boundaries of a single region are published as separate markers with the same color.

• insets ([visualization_msgs/PolygonArray])

  A set of polygons that are at a slight inward offset from boundaries. There might be more insets than boundaries since the inward shift can cause a single region to break down into multiple when narrow passages exist.

Plugins

You can create your own plugin to process the elevation map and publish as a layer in GridMap message.

Let's look at the example.

First, create your plugin file in elevation_mapping_cupy/script/plugins/ and save as example.py.

import cupy as cp
from typing import List
from .plugin_manager import PluginBase


class NameOfYourPlugin(PluginBase):
    def __init__(self, add_value:float=1.0, **kwargs):
        super().__init__()
        self.add_value = float(add_value)

    def __call__(self, elevation_map: cp.ndarray, layer_names: List[str],
            plugin_layers: cp.ndarray, plugin_layer_names: List[str])->cp.ndarray:
        # Process maps here
        # You can also use the other plugin's data through plugin_layers.
        new_elevation = elevation_map[0] + self.add_value
        return new_elevation

Then, add your plugin setting to config/plugin_config.yaml

example:                                      # Name of your filter
  type: "example"                             # Specify the name of your plugin (the name of your file name).
  enable: True                                # weather to load this plugin
  fill_nan: True                              # Fill nans to invalid cells of elevation layer.
  is_height_layer: True                       # If this is a height layer (such as elevation) or not (such as traversability)
  layer_name: "example_layer"                 # The layer name.
  extra_params:                               # This params are passed to the plugin class on initialization.
    add_value: 2.0                            # Example param
    
example_large:                                # You can apply same filter with different name.
  type: "example"                             # Specify the name of your plugin (the name of your file name).
  enable: True                                # weather to load this plugin
  fill_nan: True                              # Fill nans to invalid cells of elevation layer.
  is_height_layer: True                       # If this is a height layer (such as elevation) or not (such as traversability)
  layer_name: "example_layer_large"           # The layer name.
  extra_params:                               # This params are passed to the plugin class on initialization.
    add_value: 100.0                          # Example param with larger value.

Finally, add your layer name to publishers in config/parameters.yaml. You can create a new topic or add to existing topics.

  plugin_example: # Topic name
    layers: [ 'elevation', 'example_layer', 'example_layer_large' ]
    basic_layers: [ 'example_layer' ]
    fps: 1.0        # The plugin is called with this fps.