Overview of the iMESA algorithm operating on a toy dataset
iMESA is an incremental distributed back-end algorithm for Collaborative Simultaneous Localization and Mapping (C-SLAM). For real-world deployments, robotic teams require algorithms to compute a consistent state estimate accurately, within online runtime constraints, and with potentially limited communication. In this package we provide the original implementation of "Incremental Manifold Edge-based Separable ADMM" (iMESA) a fully distributed C-SLAM back-end algorithm that can provide a multi-robot team with accurate state estimates in real-time with only sparse pair-wise communication between robots. This algorithm is described in full in our paper "iMESA: Incremental Distributed Optimization for Collaborative Simultaneous Localization and Mapping" which can be accessed on arXiv. If you use this package please cite our paper as:
@inproceedings{mcgann_imesa_2024,
title = {{iMESA}: Incremental Distributed Optimization for Collaborative Simultaneous Localization and Mapping},
author = {D. McGann and M. Kaess},
fullauthor = {Daniel McGann and Michael Kaess},
booktitle = {Proc. Robotics: Science and Systems (RSS)},
year = 2024,
pages = {n/a}, % will be added soon
address = {Delft, {NL}}
}
Extensive evaluation on real and synthetic data demonstrates that iMESA is able to outperform comparable state-of-the-art C-SLAM back-ends. Below we have an examples of such performance on a long 5 robot dataset, and a shorter 10 robot dataset. In these animations color represents the magnitude of error for individual poses.
iMESA Compared to prior works on a long 5 robot synthetic dataset
iMESA Compared to prior works on a short 10 robot synthetic dataset
We provide the implementation of iMESA as a simple C++ library. This library provides a class IMESAAgent that can be used on-board each robot in a multi-robot team to solve their collaborative state estimate using the iMESA algorithm. Practically, this class is used much like existing incremental SLAM solvers like ISAM2 as provided by gtsam, with extensions to incorporate information from communications with other robots.
See include/imesa/imesa.h for inline documentation on this class's interface. Additionally, see our experiments repository for an example of how this library can be used. The best place for documentation on the algorithm itself is the paper discussed above!
We additionally provide a more generic interface (IncrementalSAMAgent) that can be inherited from, to implement other incremental C-SLAM solvers that use the same 2-stage communication procedure. See our implementation of DDF-SAM2 in our experiments repository for an example of how to write a custom algorithm using this interface.
iMESA uses uses the keys of variables to implicitly identify which variables are shared with other robots. The library assumes that all variable keys are represented according to the gtsam::Symbol convention and consist of a character that denotes the unique identifier of the robot that owns the variable and a integer representing that index of the variable. For example the 15th pose of robot "a" will be a15. The iMESA algorithm identifies shared variables as any variable with a key denoting that it is owned by another robot.
This library also provides functionality to handle global variables that are variables in the SLAM optimization that do not necessarily belong to any singular robot (i.e. landmarks). Such variables are marked with a # character in their variable key. Handling these variables requires some additional information to be shared during communication (keys of global vars). This functionality was not explicitly discussed in the paper above, but has been tested and users are safe to use this functionality!
The only dependency for iMESA is GTSAM. iMESA is sensitive to the version of GTSAM used. We have tested iMESA with GTSAM version 4.2.0 and recommend its use.
iMESA additionally needs some development functionality in GTSAM that unfortunately is not included in any tagged release. The details of these changes are summarized in this pull request. For convenience we provide these changes cherry-picked onto GTSAM 4.2.0 via our fork under branch 4.2.0-imesa.
As this package is only a library it is most frequently will be used as a 3rd party dependency within your project. We encourage users to use FetchContext to access and make the iMESA library available within their own project.
You can include iMESA as a dependency in your project by adding the following to your CMakeLists.txt file:
include(FetchContent)
FetchContent_Declare(
imesa
GIT_REPOSITORY https://github.com/rpl-cmu/imesa.git
GIT_TAG main
)
FetchContent_MakeAvailable(imesa)
The project in which iMESA is used will need to be build against the GTSAM version discussed above. An example of this in practice can be found in the experiments repository.
If you encounter issues with this library please fill out a bug report on github with details on how to reproduce the error!