fact.rs

fact.rs (pronounced factors) is a nonlinear least squares optimization library over factor graphs written in Rust.

It is specifically geared toward sensor fusion in robotics. It aims to be fast, easy to use, and safe. The fact.rs API takes heavy inspiration from the gtsam library.

Currently, it supports the following features

• Gauss-Newton & Levenberg-Marquadt Optimizers
• Common Lie Groups supported (SO2, SO3, SE2, SE3) with optimization in Lie Algebras
• Automatic differentiation via dual numbers
• Serialization of graphs & variables via optional serde support
• Easy conversion to rerun types for straightforward visualization

We recommend you checkout the docs for more info. For usage, simply add factrs to your Cargo.toml and start using it!

Examples

There's a number of examples found in the examples folder, including loading g20 files, serialization, and custom factors.

To run a simple pose graph optimization, simply clone this repository and run,

cargo run --release --example g20 ./examples/data/M3500.g20

to visualize the optimization steps with rerun simply add --features rerun to the above command.

Running the other examples can be done similarly,

cargo run --release --example gps
cargo run --release --example serde --features serde

Additionally, we recommend checking out the tests folder for more examples of how to make custom noise models, residuals, robust kernels, and variables.

Full Example

use factrs::{
    assign_symbols,
    core::{BetweenResidual, GaussNewton, Graph, Huber, PriorResidual, Values, SO2},
    fac,
    traits::*,
};

// Assign symbols to variable types
assign_symbols!(X: SO2);

fn main() {
    // Make all the values
    let mut values = Values::new();

    let x = SO2::from_theta(1.0);
    let y = SO2::from_theta(2.0);
    values.insert(X(0), SO2::identity());
    values.insert(X(1), SO2::identity());

    // Make the factors & insert into graph
    let mut graph = Graph::new();
    let res = PriorResidual::new(x.clone());
    let factor = fac![res, X(0)];
    graph.add_factor(factor);

    let res = BetweenResidual::new(y.minus(&x));
    let factor = fac![res, (X(0), X(1)), 0.1 as std, Huber::default()];
    graph.add_factor(factor);

    // Optimize!
    let mut opt: GaussNewton = GaussNewton::new(graph);
    let result = opt.optimize(values).unwrap();
    println!("Results {:#}", result);
}

Compile-time Errors

fact.rs leans into the Rust way of doing things, and attempts to compile-time error as much as possible. This includes the following,

• Symbols are assigned to variables at compile-time, ensuring that symbols are can not be mismatched
• Enforcing the correct number of keys for a factor
• Ensuring that noise model dimensions match the residual dimensions

A few examples,

use factrs::core::{assign_symbols, fac, PriorResidual, Values, VectorVar2, SO2};

// Assign symbols to variable types
assign_symbols(X: SO2, Y: SO2);
let mut values = Values::new();

// Proper usage
let id = SO2::identity();
values.insert(X(0), id);
let prior = PriorResidual::new(id);
let f = fac![prior, X(0), (0.1, 0.2) as std];

// These will all compile-time error
// mismatched symbol-variable types
values.insert(X(5), VectorVar2::identity());
// wrong number of keys 
let f = fac![PriorResidual::new(id), (X(0), X(1))]; 
// wrong noise-model dimension
let n = GaussianNoise::<5>::from_scalar_sigma(0.1);
let f = fac![PriorResidual::new(id), X(0), n];
// mismatched symbol-variable types
let f = fac![PriorResidual::new(id), Y(0), 0.1 as std];

Benchmarks

Performance-wise, factrs is competitive with alternative libraries. Benchmarks were ran on a 12th Gen Intel i9 and are all single-threaded (for now). Current benchmarks include gtsam and tiny-solver-rs and data can be found in the examples/data folder.

2D Benchmarks

benchmark	args	fastest	median	mean
factrs	M3500	81.23 ms	82.13 ms	82.80 ms
gtsam	M3500	160.00 ms	161.13 ms	161.14 ms
tiny-solver	M3500	125.13 ms	130.46 ms	132.08 ms

3D Benchmarks

benchmark	args	fastest	median	mean
factrs	sphere2500	352.97 ms	355.01 ms	355.14 ms
gtsam	sphere2500	389.81 ms	395.16 ms	396.65 ms
tiny-solver	sphere2500	600.80 ms	615.90 ms	616.75 ms
factrs	parking-garage	292.14 ms	294.01 ms	294.10 ms
gtsam	parking-garage	113.24 ms	114.74 ms	114.45 ms
tiny-solver	parking-garage	329.48 ms	334.78 ms	335.17 ms

Note, gtsam is significantly faster for the parking garage due to leveraging the sparsity of the pose graph better using the Baye's tree, something that is planned for factrs.

To run the rust benchmarks after cloning, simply run,

cargo bench -p factrs-bench

and the C++ benchmarks can be run with,

cmake -B build factrs-bench/cpp
cmake --build build
./build/bench

both of which have alias commands in the root justfile.

There is still some benchmarking work to be done and we'd love some help if you'd like an easy way to contribute! There's a few libraries that could be added, specifically ceres and sophus-rs. Additionally, it'd be nice if all benchmarks had a rust frontend using FFI for easier running - this was begun in the easton/benches branch.

Installation

Simply add via cargo as you do any rust dependency,

cargo add factrs

Contributions

Contributions are more than welcome! Feel free to open an issue or a pull request with any ideas, bugs, features, etc you might have or want.

We feel rust and robotics are a good match and want to see rust robotics libraries catch-up to their C++ counterparts.