kornia-slam

The v0.1.0 binary predates run files: it takes a source subcommand and flags (kornia-slam --help, kornia-slam euroc --help) rather than the --config run files described below. Global options go before the subcommand, e.g. loop closing with the vocabulary from Quick start: kornia-slam --vocab /path/to/ORBvoc.txt --apply-pgo euroc ….

It is built with the default features: Rerun streaming (--rerun-stream, which needs the rerun 0.29.2 viewer on PATH, e.g. pip install rerun-sdk==0.29.2) is in, live OAK-D and UVC cameras are not. Build from source with --features oakd or --features uvc for those; OAK-D has extra prerequisites, see apps/kornia-slam-app.

From source

The app needs Rust 1.91+ (the libraries 1.89+). Build and run with Cargo as in Quick start, or with Pixi, which provides the toolchain:

pixi run rust-run -- --config configs/euroc.ron
pixi run rust-lint
pixi run rust-test

Quick start

Download a EuRoC sequence (ASL format) to data/euroc/MH_01_easy, or edit the data path in the run file, then:

# monocular
cargo run --release -p kornia-slam-app -- --config configs/euroc.ron

# stereo + IMU with loop closing, with evaluation against ground truth
cargo run --release -p kornia-slam-app -- --config configs/euroc-stereo-imu-loop.ron --evaluate

Loop closing needs ORB-SLAM3's vocabulary: extract Vocabulary/ORBvoc.txt.tar.gz from ORB-SLAM3 to weights/ORBvoc.txt.

A run file names the source and the system to run on it; see apps/kornia-slam-app.

More sources and options: apps/kornia-slam-app.

Library

A system configuration and the source's calibrated rig build a SlamSystem, which then takes IMU samples and images as they arrive and returns poses. Feature extraction, stereo matching and frame history are the system's; sources supply synchronized, calibrated (and, for stereo, rectified) images.

use kornia_slam::{CameraSelection, SensorSelection, SlamSystem, SystemConfig};

let config = SystemConfig {
    sensors: SensorSelection { cameras: CameraSelection::Stereo, imu: true },
    ..SystemConfig::default()
};
let mut system = SlamSystem::build(config, rig)?; // rig: kornia_slam::SensorRig
for (idx, frame) in frames.enumerate() {
    system.process_imu(&frame.imu);
    let result = system.process_stereo(frame.timestamp_sec, &frame.left, &frame.right)?;
    println!("{idx}: {:?} {:?}", result.status, result.pose_world_to_cam);
}

With the serde feature, SystemConfig::from_ron_file reads the same settings from a RON file, the system section of a run file.

Integrations

• Copper: stereo visual-inertial odometry as Copper tasks via cu-kornia-vio. See INTEGRATIONS.md for setup, graph wiring and timing requirements.

Development

With Pixi, which sets up the toolchain for you:

pixi run rust-lint           # fmt + clippy + check
pixi run rust-test           # workspace tests
pixi shell                   # drop into the environment

Or with a Rust toolchain (1.91+; the library crates alone need 1.89) installed yourself:

cargo fmt --all -- --check
cargo clippy --all-targets -- -D warnings
cargo test --workspace

Paper

Kornia-SLAM: An End-to-End Visual-Inertial SLAM System in Rust
Christie J. Purackal, Edgar Riba, Astik Srivastava.
Rust for Robotics workshop at IROS 2026 (preprint).
Paper (PDF) · Poster (PDF)

The paper's EuRoC results come from the August 2026 code, before v0.1.0. To cite kornia-slam, use GitHub's "Cite this repository" button, which reads CITATION.cff.

License

Apache-2.0