Voxgraph is a globally consistent volumetric mapping framework. It represents the world as a collection of Signed Distance Function submaps, which it aligns through pose graph optimization. Each submap is based on voxblox and the collection itself is managed using c-blox.
Local consistency is maintained through registration constraints between overlapping submap pairs. The computational cost of these constraints is significantly reduced by directly exploiting the underlying Signed Distance Function representation of the submaps. This makes it possible to globally optimize maps with hundreds of overlapping submap pair constraints, in a matter of seconds, even on computationally constrained platforms.
Global consistency can be achieved by adding absolute pose or loop closure constraints to the pose graph. We expect to provide an internal mechanism in voxgraph to detect global loop closures in the future. In the meantime, an interface to add loop closures from external sources is provided through a ROS topic. Absolute pose constraints are mostly implemented, but this work has not yet been completed since we don't have a direct use-case for them. Please create a GitHub issue if you would like to use them. Of course, pull requests are even more welcome.
When using voxgraph for research, please cite the following paper [preprint | IEEE]:
@ARTICLE{reijgwart2020voxgraph,
author={V. {Reijgwart} and A. {Millane} and H. {Oleynikova} and R. {Siegwart} and C. {Cadena} and J. {Nieto}},
journal={IEEE Robotics and Automation Letters},
title={Voxgraph: Globally Consistent, Volumetric Mapping Using Signed Distance Function Submaps},
year={2020}}
Abstract
Globally consistent dense maps are a key requirement for long-term robot navigation in complex environments. While previous works have addressed the challenges of dense mapping and global consistency, most require more computational resources than may be available on-board small robots. We propose a framework that creates globally consistent volumetric maps on a CPU and is lightweight enough to run on computationally constrained platforms. Our approach represents the environment as a collection of overlapping signed distance function (SDF) submaps and maintains global consistency by computing an optimal alignment of the submap collection. By exploiting the underlying SDF representation, we generate correspondence-free constraints between submap pairs that are computationally efficient enough to optimize the global problem each time a new submap is added. We deploy the proposed system on a hexacopter micro aerial vehicle (MAV) with an Intel i7-8650 U CPU in two realistic scenarios: mapping a large-scale area using a 3D LiDAR and mapping an industrial space using an RGB-D camera. In the large-scale outdoor experiments, the system optimizes a 120 × 80 m map in less than 4 s and produces absolute trajectory RMSEs of less than 1 m over 400 m trajectories. Our complete system, called voxgraph, is available as open-source.
Voxgraph extends voxblox and runs on ROS. Therefore start by installing and configuring these as described in the voxblox installation instructions.
Next, clone this repository and its dependencies into the same catkin workspace. If you have SSH keys for GitHub set up:
cd ~/catkin_ws/src
git clone [email protected]:ethz-asl/voxgraph.git
wstool merge ./voxgraph/voxgraph_ssh.rosinstall
wstool update
No ssh keys?
cd ~/catkin_ws/src/
git clone https://github.com/ethz-asl/voxgraph.git
wstool merge ./voxgraph/voxgraph_https.rosinstall
wstool update
Then compile voxgraph and its dependencies with:
cd ~/catkin_ws/src/
catkin build voxgraph
We provide a demo dataset here, which features a hexacopter flying through an indoor-outdoor search and rescue training site. The rosbag includes the robot's visual, inertial and LiDAR sensor data, and the GPS RTK measurements used for the evaluations in the paper.
Running this example requires a couple more packages than the basic install, for performing lidar undistortion. Clone and build these using:
cd ~/catkin_ws/src/
wstool merge ./voxgraph/arche_ssh.rosinstall
wstool update
catkin build lidar_undistortion
No ssh keys?
cd ~/catkin_ws/src/
wstool merge ./voxgraph/arche_https.rosinstall
wstool update
catkin build lidar_undistortion
Download and unzip the dataset
wget -P ~/Downloads http://robotics.ethz.ch/~asl-datasets/2020_voxgraph_arche/arche_flight1_2ms_indoor-outdoor-figure-8.zip
unzip ~/Downloads/arche_flight1_2ms_indoor-outdoor-figure-8.zip
Launch the example
roslaunch voxgraph arche_demo.launch rosbag_path:=${HOME}/Downloads/arche_flight1_2ms_indoor-outdoor-figure-8.bag
If everything has worked correctly, this should open rviz and within a few seconds the first fragment of the map should appear.
Note that visualizing the whole mesh can be heavy, particularly as the map grows. If you are on a weak computer, consider disabling the mesh visualization once you have confirmed voxgraph is running correctly (uncheck the box next to "Merged Map" in rviz). The current state of the map can be saved at any time using a service call:
rosservice call /voxgraph_mapper/save_combined_mesh "file_path: '$HOME/mesh.ply'"
This mesh is viewable through a number of programs. For example using meshlab:
meshlab ~/mesh.ply
The basic requirements for running voxgraph are:
- an odometry source, and
- a source of dense (depth) data.
The system is agnostic as to where the odometry comes from. In the demo above we used visual odometry provided by rovio. However, we have also obtained good results using other odometry sources such as LOAM.
The requirement is that there is a link on the tf tree between the odometry frame and the robot's base_link frame. The names for the frames can be configured using the input_odom_frame
and input_base_link_frame
rosparams. In the demo, we for example set them here.
For depth data we have tried RGB-D and 3D LiDAR sensors. The input datastream is provided to the system as rostopic, whose name can be configured through the pointcloud_topic
rosparam. See here for the topic used in the above example.
The frame convention used by voxgraph is shown on the right. The orange arrow corresponds to the odometry input. The black arrows correspond to the transforms that are being published by voxgraph. The names of all these frames can be configured using these rosparams.
Setting the output_odom_frame
to the same name as the input_odom_frame
, and the output_base_link_frame
to the same name as the input_base_link_frame
, would give you a single connected tf tree as shown in the diagram.
If you would rather keep the input and output tf trees disjoint, you could set different names for the output_odom_frame
and/or output_base_link_frame
. In this case voxgraph would automatically republish the odometry, using the new frame names you chose, each time it receives a pointcloud.