Auto-tuned ICP

Auto-tuned Iterative Closest Point (AICP) is a module for laser-based localization and mapping (Nobili et al., ICRA 2017). The implementation of AICP includes a module for localization failure prediction (Nobili et al., ICRA 2018). The registration strategy is based on the libpointmatcher framework (Pomerleau et al., AR 2012).

Demonstration Video with ANYMal:

AICP has been tested on:

• Carnegie Robotics Multisense SL data from:
  • NASA Valkyrie humanoid
  • Boston Dynamics Atlas humanoid
  • IIT HyQ quadruped
  • Clearpath Husky mobile platform
• Velodyne HDL-64 data from
  • KITTI dataset (mobile platform)
• Velodyne VLP-16 data from
  • ANYbotics ANYmal quadruped

For inter-process communication, the package includes wrappers for:

• ROS
• LCM (discontinued)

Description

The core AICP strategy is compiled in aicp_core. The default steps perform frame-to-reference localization and mapping, and include:

1. On a thread, AICP accumulates laser scans, i.e. each 3D point cloud processed for registration is constituted of batch_size scans
2. On a second thread, it stores the first point cloud as the reference cloud
3. Prior to registration, AICP computes an octree-based overlap parameter (Nobili et al., ICRA 2018) between the current and reference point clouds, and uses it to auto-tune online the outlier rejection filter of the registration strategy
4. A reference point cloud update can be trigger either in a windowed fashion (reference_update_frequency) or based on the risk of alignment failure prediction (failure_prediction_mode)
5. AICP registers each new point cloud to the reference point cloud
6. Finally, AICP publishes a corrected body pose estimate at the frequency of the state estimator

Additional Functionalities

• Localization only mode -- This localizes against a fixed input map

Quick Start

Dependencies

TODO: add complete list of apt install packages

sudo apt-get install libpointmatcher-dev libpointmatcher1

How To Compile

The code is cloned and built in the standard manner for a ROS package.

Tips:

• We currently support ROS Melodic / Ubuntu 18.04. The code is also compatiable with ROS Kinetic / Ubuntu 16.04.
• Configure your workspace to be in Release mode:

catkin config --cmake-args -DCMAKE_BUILD_TYPE=Release

How To Run

ROS

• User Interface:
  1. roscore
  2. EITHER rosrun rviz rviz -d [PATH_TO]/drs_base/git/aicp/aicp_ros/config/rviz/simple_aicp.rviz
  3. OR roslaunch aicp_ros view_recorded_rosbag_aicp_simple.launch
• AICP launch:
  • AICP mapping -- frame-to-reference: roslaunch aicp_ros aicp_mapping.launch
  • AICP localization only -- frame-to-map: roslaunch aicp_ros aicp_localization_only.launch
• Log:
  • rosbag play --clock --pause yourbagfile.bag
  • You can download the bag file from the experiment in the video here:
    • ftp://ftp.robots.ox.ac.uk/pub/outgoing/aicp/anymal_2019-04-17-12-59-42_filtered_aicp_input.bag

LCM (discontinued)

Help: rosrun aicp_lcm aicp_lcm_node -h Example: rosrun aicp_lcm aicp_lcm_node -s debug -b 80 -f 5

Note:

• Option "-s robot" is meant to be used if the corrected pose from AICP is fed back into the state estimator. Each corrected pose is published only once as soon as computed.
• Option "-s debug" is meant to be used during debug. The corrected pose message is published at the same frequency as the pose estimate from the state estimator (for visualization purposes).

Misc

Required Topics (ROS):

• Point Cloud from a 3D Lidar, ideally transformed into the base frame
• Relative odometry estimate. This is the estimate of the LIDAR frame in a fixed frame. This should have low drift

To filter an ANYmal rosbag to contain only the required topics

• rosbag filter in.bag out.bag "topic=='/tf' or topic=='/state_estimator/pose_in_odom' or topic=='/point_cloud_filter/velodyne/point_cloud_filtered' or topic=='/tf_static'"

Main dependencies

• libpointmatcher, a modular library implementing the Iterative Closest Point (ICP) algorithm for aligning point clouds.

• Point Cloud Library (PCL) revision pcl-1.7.0, a standalone, large scale, open project for 2D/3D image and point cloud processing.

• Octomap, a probabilistic 3D mapping framework based on octrees.

• OpenCV -- Open Source Computer Vision Library.

• Eigen, a C++ template library for linear algebra: matrices, vectors, numerical solvers, and related algorithms.

Credits

@inproceedings{Nobili17icra,
  title = {Overlap-based {ICP} Tuning for Robust Localization of a Humanoid Robot},
  author = {S. Nobili and R. Scona and M. Caravagna and M. Fallon},
  booktitle = {{IEEE International Conference on Robotics and Automation (ICRA)}},
  location = {Singapore},
  month = {May},
  year = {2017},
}

@inproceedings{NobiliCamurri17rss,
  title = {Heterogeneous Sensor Fusion for Accurate State Estimation of Dynamic Legged Robots},
  author = {S. Nobili and M. Camurri and V. Barasuol and M. Focchi and D. Caldwell and C. Semini and M. Fallon},
  booktitle = {{Robotics: Science and Systems (RSS)}},
  location = {Boston, MA, USA},
  month = {July},
  year = {2017},
}

@inproceedings{Nobili18icra,
  title = {Predicting Alignment Risk to Prevent Localization Failure},
  author = {S. Nobili and G. Tinchev and M. Fallon},
  booktitle = {{IEEE International Conference on Robotics and Automation (ICRA)}},
  location = {Brisbane, Australia},
  month = {May},
  year = {2018},
}

License

The License information is available in the LICENSE file contained in this project repository.

Author: Simona Nobili, March 2019. Email: snobili@robots.ox.ac.uk, simona.nobili@ed.ac.uk