Rerun

🚀 Introducing Kornia-SLAM v0.1.0 We’re excited to release Kornia-SLAM, an open-source, real-time visual-inertial SLAM system written in Rust and built on top of kornia-rs. Kornia-SLAM is an effort to explore what a modern, modular SLAM stack can look like in the Rust ecosystem, bringing together classical geometric vision, state estimation, optimisation, and robotics tooling in a safe and performant systems language. The first release includes: 🔹 Monocular, stereo, and visual-inertial ORB SLAM 🔹 Bundle adjustment and Pose-graph optimisation 🔹 DBoW2 place recognition and loop closure 🔹 EuRoC, MCAP, OAK-D, and webcam inputs 🔹 Rerun visualisation and a TUI mode built with Ratatui 🔹 Integration with the Copper robotics runtime Copper Robotics Inc. Kornia-SLAM was introduced at the Rust for Robotics Workshop at IROS 2026 as part of Edgar Riba’s talk. This is still an early-stage project, and there is plenty we want to build next. We’d love contributions from the computer vision, SLAM, robotics, and Rust communities, whether that’s new sensors, algorithms, optimisations, benchmarks, or integrations.⭐ Check out the project, try it out, and let us know what you think: https://lnkd.in/e6WeHS8t www.kornia.org/slam/ Christie Purackal · Edgar Riba · Astik Srivastava #Kornia #SLAM #Robotics #ComputerVision #RustLang #VisualInertialOdometry #OpenSource #IROS2026

Here it is! I rebuilt a visual iniertial multicamera pipeline in rust + python, alll with support for Rerun OSS data catalog + visualization. The core is rust to make sure it runs fast, but you can also easily call it from python. I've been unhappy with pretty much all of the other vio pipelines out there, as none of the ones I know about hit the requirements I wanted 1. Multi Camera - I'm convinced 4 or more cameras are optimal here, sure you can do purely monocular or stereo, but IMO most robots and actual deployed systems out there that need to run at realtime on low powered hardware should AND need to be robust against occlusion/motion blur/ect need at least a stereo pair. Theres a reason aria gen2 switched to 4 global shutter cameras. The cuvslam paper argues the same point for general robotics. 2. IMU support - bascially the same point as above, realtime robotics needs to be robust, and imu + cameras are perfect for each other 3. Diverse hardware - I didn't want this to just run on pytorch/nvidia devices. I want this to work on my linux machine, or my mac mini, or my Raspberry Pi! 4. Hardware acceleration - In the same way I want diversity of hardware, I also want to make sure we can take advantage of hardware acceleration. If I have a 5090, I should get the most out of it =] and if I have a rpi5 I should also get the most out of it! It has a GPU on it after all 5. Opensource - I'm sure theres plenty of great vio/slam pipelines out there that do the above, but none that are open that I know of =/ So thats exactly what I did, taking inspiration from basalt/cuvslam/glide this repo has imu+multicamera support, works on diverse hardware (tested on my mac mini + 5090 linux machine + rpi5 + rockchip 3588) AND support GPU thanks to the awesome CubeCL library that lets you write kernels on rust and support wgpu/metal/vulkan I also leaned heavily on kornia_foss and the great work done by the folks there. I used many of the components in kornia-rs and got lots of inspiration from kornia-slam I plan to rip out parts that make sense and contribute them back to kornia. The biggest problem now is that this is VIO, so it has lots of drift. I tried walking for a mile or two and returning to the same spot and the drift is pretty bad. That'll a problem for later me, but shouldn't be too hard to add loop closure. I'm looking at cuSFM + colmap for this.

As part of my exo ego pipeline, I haven't yet found a VIO/Slam solution that hits all of the features I want. Now that agents make it super easy to write low level systems code, I've almost finished building a portable rust/python slam pipeline that works on both a 5090 AND a rockchip. To really make this work I needed a good ground truth, and the wonderful work from Collabora and Mateo De Mayo on the monado slam dataset has been indispensable. I registered it as a Rerun dataset in the OSS catalog to make it very easy to verify that the GT actually looks right and use it as an eval, and it has worked REALLY well. I'll have more on the VIO pipeline here in the coming days

Inspired by R2-D2, I’m building Pollen Robotics's Reachy Mini into a robot whose perception, memory, and movement work together as one coordinated system. The navigation system combines live RGB-D mapping, object detection, persistent spatial memory, path planning, and person-following. Reachy searches for a user hierarchically: scanning with its head, rotating its body, then repositioning on its wheels. Once it finds someone, it can turn, approach, and maintain a safe following distance while visualizing its perception and 3D map through Rerun. Much of the challenge has been reliability outside a clean demo: changing coordinate frames, intermittent Wi-Fi, stale sensor data, tracking loss, motor safety, and physical constraints. There’s plenty left to improve, but it’s beginning to feel like a real robotic system. #Robotics #ComputerVision #RGBD #PathPlanning #EmbeddedSystems #ReachyMini