Rerun 0.11 brings all SDKs to parity

Rerun helps engineers and researchers log and visualize streams of multimodal data. In fields like robotics and augmented reality, that data often lives in programs written in C++. In Rerun 0.10, we released our first version of the Rerun C++ SDK, and with 0.11 it is now fully on par with our two other SDKs in Python and Rust.

Thanks to some great community feedback we've added more options for integrating Rerun in your CMake projects. Rerun 0.11 comes with big improvements to both the performance and ergonomics of logging non-Rerun data. It also includes multiple general performance improvements that bring the C++ SDK on par with the Rust version.

While both performance and nice API's are great, the final and probably most important piece to the C++ puzzle is that 0.11 brings hosted API reference documentation.

In addition to C++ improvements, this version also comes with more control over the time ranges used for e.g. time series plots. We've also deployed the web viewer as a stand alone npm package for easier integration in other tools.

Check out the release notes for a full list of changes.

Full time range queries

Rerun has long had a feature called "Visual History", where you could choose to include previous data for an entity shown in a view. For instance, you could use that to accumulate point clouds that have been observed over time in a 3D scene. In 0.11, we are expanding that feature to a full "Visible Time Range" query where you'll be able to set both the start end end of the time range to include.

An important use case that the "Visible Time Range" feature enables is to show windowed time series plots, which is important for live streaming visualizations. This is the first step in what will be a series of improvements to Rerun's plotting capabilities.

You can try out the new "Visible Time Range" feature right now in your browser at

Using Rerun with OpenCV and Eigen in C++

We've put together a minimal C++ example of using Rerun together with OpenCV and Eigen that should hopefully be a good starting off point to both see how to use Rerun with those popular libraries, as well as how you can use Rerun with your own types.

Check out the example here.

Rerun Web Viewer on npm

To make it easier to use Rerun for viewing data on the web, we've published the viewer as a stand alone npm package. For those of you building in React, we also published a React component that you can use directly.

The viewer is super easy to integrate in your web app but you currently only have limited control over options and layout from javascript.

import { WebViewer } from "@rerun-io/web-viewer"; // Either pass the url of an .rrd file const DATA_SOURCE_URL = ""; // Or pass the websocket url to a live Rerun stream // const DATA_SOURCE_URL = "ws://localhost:9877"; const parentElement = document.body; const viewer = new WebViewer(); await viewer.start(DATA_SOURCE_URL, parentElement); // … viewer.stop();

The data that you pass to the Viewer can either come from a websocket connection to the SDK opened via the serve API, or from a hosted .rrd file.

CollectionAdapter gives better logging of non-Rerun types

Rerun 0.11 sunsets ComponentBatchAdaptor in favor of the new CollectionAdapter, which is more expressive and general. The end result is that the code to get data in your own formats logged to Rerun will be simpler and produce fewer unnecessary copies.

The ComponentBatchAdaptor made it easy to map your own types to Rerun's ComponentBatches. For example, if you have a proprietary Vec3D and represent point clouds as std::vector<Vec3D>, you could use the ComponentBatchAdaptor to log it as a Rerun batch of rerun::Position3Ds without incurring extra copies.

However, single images do not fit this pattern and would therefore require an intermediate copy to log. The new CollectionAdapter fixes this in 0.11 by generalizing to inner types in addition to batches.

For example, to log an OpenCV cv::Mat to Rerun as an image, you might write the following adapters:

// Adapters so we can easily borrow an OpenCV image into Rerun images without copying: template <> struct rerun::CollectionAdapter<uint8_t, cv::Mat> { Collection<uint8_t> operator()(const cv::Mat& img) { assert(CV_MAT_DEPTH(img.type()) == CV_8U); return Collection<uint8_t>::borrow(, * img.channels()); }; }; // Convenience function for extracting the image shape. rerun::Collection<rerun::TensorDimension> tensor_shape(const cv::Mat& img) { return { static_cast<size_t>(img.rows), static_cast<size_t>(img.cols), static_cast<size_t>(img.channels()), }; };

And then log it to Rerun like:

cv::cvtColor(img, img, cv::COLOR_BGR2RGB); // Rerun expects RGB format rec.log("image", rerun::Image(tensor_shape(img), rerun::TensorBuffer::u8(img)));

For images, we've also used the CollectionAdapter to implement a built-in convenience constructor so that you can log your own images to Rerun without having to write an adapter for the tensor buffer:

// Creates an image archetype from a shape and pointer rec.log("image", rerun::Image(tensor_shape(img),;

Your feedback impacts what we build

The early feedback we got on 0.10 had a big impact on the C++ improvements we shipped in 0.11. We love hearing everything about what's worked for you and how we can improve things. If it's bugging you we'd love to know, even if it seems minor!

Join us on Github or Discord and let us know what you like and what you'd hope to see change in the future.

Introducing the Rerun SDK for C++

The ability to log streams of multimodal data from C++ and visualize it live with Rerun has been our most requested feature since before the public launch in February. The C++ SDK is finally out, but getting here the right way has been a long road.

If you're eager to get started, the quick start guide is right here.

We designed the C++ API's to be easy to use, efficient, and consistent with our API's in Rust and Python. Another key goal was to make it easy to write adapters to log data in your own custom formats to Rerun. With the first iteration of our C++ API released, we're really looking forward to hear from the community on how well that works, and what we can do to improve it further.

Here is an example of creating and logging a random tensor in Python, Rust, and C++:

"""Create and log a tensor.""" import numpy as np import rerun as rr tensor = np.random.randint(0, 256, (8, 6, 3, 5), dtype=np.uint8) # 4-dimensional tensor rr.init("rerun_example_tensor_simple", spawn=True) # Log the tensor, assigning names to each dimension rr.log("tensor", rr.Tensor(tensor, dim_names=("width", "height", "channel", "batch")))

Which should look like this for all languages (ignoring different random numbers):

You can integrate Rerun into your CMake-based project by adding the following snippet to your CmakeLists.txt file:

include(FetchContent) FetchContent_Declare(rerun_sdk URL FetchContent_MakeAvailable(rerun_sdk)

For more information on working with the C++ APIs, see the C++ quick start or the logging tutorial.

While this release is all about C++, Rerun 0.10 also includes an in-app getting started guides using Rerun's markdown support.

As a reminder, Rerun is under active and high paced development. While the C++ SDK is feature complete, i.e. everything you can log in Python and Rust can be logged in C++, several improvements to the C++ SDK are already planned for 0.11.

The road to C++

Rerun is currently used by researchers and engineers in fields like computer vision, robotics, and AR/XR. The norm in these fields is that running in production (on the edge) means C++. It's therefore no surprise that C++ bindings have been our most requested feature. We've still held off releasing them until now.

Held back by complexity

Within the first few months of building Rerun in earnest, it became clear that manually maintaining nice, well documented, and consistent APIs between just Python and Rust was very hard. Too many forces naturally pulled them apart. If we didn't make some major changes first, adding C++, which is much harder to work with than Rust, would slow us to a crawl.

Most of the complexity arises from that the Rerun SDKs lets users log many kinds of data, which could be represented in many different formats in their code. Rerun uses Apache Arrow for in-memory data and Arrow IPC for in-flight data. The SDKs need to make it super easy to get user data for all those types into Rerun's Arrow based format, while both feeling consistent across languages and feeling native to each language separately.

The solution was a new code generation framework

Five months ago, we made the decision that the only sustainable way forward was code generation. We built a system that took Rerun types defined in a new Interface Definition Language (IDL), and used them to generate working SDK code in Rust and Python. The framework is entirely written in Rust, uses the Rust build system as a build graph for the whole generation process, and probably deserves a post or two on its own.

In total it took about four months get the code generation framework in place, move all SDK and engine code onto it, and move over to the more type oriented API we released in Rerun 0.9. To make sure it would work for new languages, we tested it on parts of the C++ SDK in parallel.

Since 0.9 we've used the new framework to quickly bring C++ to feature parity, and make sure that the new SDK produces 100% compatible Arrow payloads with the two existing APIs.

Cross language equivalence testing

An added benefit of the architecture where IDL specified type definitions produce SDK code, that in turn produce Arrow IPC streams, is that it makes it easy to compare outputs across languages. We built a tool that compares multiple .rrd files (Rerun's file format holding the Arrow IPC streams). That tool is used as part of a small framework that compares the output of code snippets in Python, C++, and Rust, and makes sure they all produce the exact same in-memory Arrow representation.

We use this cross-language equivalence testing framework to test all the SDKs, including the code examples used for documentation. Having these tests set up made all the difference in being able to deliver the C++ SDK so quickly after the big 0.9 release.

Making third party types loggable with ComponentBatchAdaptor

When using your own types with Rerun in C++, you have the option to define completely custom archetypes and components just like in Python and C++. However, we also want to make is as easy and performant as possible to use your own data types together with Rerun's built-in types.

For example, say you represent your point clouds as std::vector<Eigen:Vector3f>. You would then like to be able to log a point cloud to Rerun like:

std::vector<Eigen:Vector3f> points3d_vector = ... rec.log("points", rerun::Points3D(points3d_vector));

We can do this with ComponentBatchAdaptor. Here is how it could look for the example above:

// Adapters so we can log Eigen vectors as Rerun positions: template <> struct rerun::ComponentBatchAdapter<rerun::Position3D, std::vector<Eigen::Vector3f>> { ComponentBatch<rerun::Position3D> operator()(const std::vector<Eigen::Vector3f>& container) { return ComponentBatch<rerun::Position3D>::borrow(, container.size()); } ComponentBatch<rerun::Position3D> operator()(std::vector<Eigen::Vector3f>&& container) { throw std::runtime_error("Not implemented for temporaries"); } };

If you have your own types laid out similarly enough to the matching Rerun type, logging can usually be done without incurring any extra copies. Check out this basic example to see how to use Rerun with Eigen and OpenCV, and learn more about using ComponentBatchAdapter here.

Key current limitations

ComponentBatchAdaptor is not as expressive and performant as is should be

You can't currently write adaptors for single component batches like images or tensors. This means logging these types are both unergonomic and produce an unnecessary copy. The adapters aren't currently able to model data with strides, which means logging strided data also requires an extra copy. We will work on these shortcomings over the next few releases.

No built-in adapters for common libraries like Eigen and OpenCV

We plan to add built-in adapters for common types in Eigen and OpenCV over the next releases. If there are other libraries you think should have built-in adapters, we're happy for both suggestions and pull requests.

No hosted API documentation

While the general documentation treats C++ as a first class citizen, we don't yet have hosted API documentation like we do for Python and Rust. We should have this resolved within the next few releases but in the meantime you can find the API documentation directly in the header files. There are also C++ code examples for all our types at

Rerun only supports C++17 and later

If you need C++14 or C++11 support, please let us know.

Try it out and let us know what you think

We're really looking forward to hear how this first iteration of our C++ SDK works for you. Start with the C++ quick start guide and then join us on Github or Discord and let us know what you like and what you'd hope to see change in the future.

Also, huge thanks to everyone who tried out early builds and came with feedback, suggestions, and bug reports!

Rerun 0.9 gives access to the underlying ECS

We’re building a fast and easy to use general framework for handling and visualizing streams of multimodal data. This is a big undertaking, and the way we’re getting there is by starting with a fast and easy to use visualizer for computer vision, and then making it more capable and extensible piece by piece.

Rerun 0.9.0 is released two months after 0.8.0, but it’s been even longer coming. It includes the foundation of the coming C++ SDK, our most asked for feature by far. In order to maintain great and consistent APIs across Rust, Python, C++, and any future languages, we’ve rebuilt much of Rerun's data infrastructure around a new code generation framework.

0.9 also adds support for logging markdown, a new in-viewer getting started experience, and as always, a bunch of performance improvements.

Load example recordings, including descriptions of how they were made, directly in the viewer.

This has all been a huge lift, but what I’m most excited about are our redesigned APIs and what they pave the way for in future releases. From this release on, we’ll start to expose more and more of Rerun's underlying infrastructure, starting with the core data model, a hierarchical and time varying Entity Component System (ECS).

To ease the transition for Python users, we've marked the old APIs as deprecated in 0.9 with migration instructions in the warning messages. The old API will be removed completely in 0.10. Check out the migration guide for more details on updating your code.

A more type centric logging API

At the heart of Rerun is the ability to handle streams of multimodal data, e.g. images, tensors, point clouds, and text. To get data out of your programs and ready to be visualized, you log it with the Rerun SDK. Rerun handles everything needed to make that work. It doesn't matter if the data source and visualization are in the same process or the data is coming in real-time from multiple devices.

The ease of use, expressiveness, and extensibility of these APIs are core to the usefulness of Rerun. On a first glance, the API changes introduced in 0.9 are very small. For example, here is how you might log a single colored point cloud, represented by two 3xN numpy arrays of positions and colors.

Old Python API

rr.log_points("example/points", positions, colors=colors)

New Python API

rr.log("example/points", rr.Points3D(positions, colors=colors))

Both these log calls take the same user data. The difference is in the data type information, which is moved from the function name rr.log_points to a type, rr.Points3D that wraps the logged data. This new structure both opens up for more direct control of the underlying ECS and for more ergonomic logging of your own objects.

Lower level control of Entity Components

Rerun comes with a set of built in archetypes like rr.Points3D , rr.Image, and rr.Tensor. An archetype defines a bundle of component batches that the Rerun Viewer knows how to interpret, such that Rerun will just do the right thing™️ when you log it. In this case, that’s one component batch for positions and one for colors.

rr.log("example/points", rr.Points3D(positions, colors=colors)) # is equivalent to rr.log("example/points", rr.Points3D(positions, colors=colors).as_component_batches()) # which in this case is the same as rr.log("example/points", [rr.Points3D.indicator(), rr.components.Position3DBatch(positions), rr.components.ColorBatch(rgb=colors)])

Partial updates using the component level API

In most cases, you’ll want to stick to the high level archetype API, but directly setting single components gives a lot of control, which can matter. For instance, a common use case is meshes where only the vertex positions change over time. Logging the whole mesh for each change adds a lot of overhead. For example:

import numpy as np import rerun as rr # pip install rerun-sdk rr.init("rerun_example_mesh3d_partial_updates", spawn=True) vertex_positions = np.array([[-1.0, 0.0, 0.0], [1.0, 0.0, 0.0], [0.0, 1.0, 0.0]], dtype=np.float32) # Log the initial state of our triangle rr.set_time_sequence("frame", 0) rr.log( "triangle", rr.Mesh3D( vertex_positions=vertex_positions, vertex_normals=[0.0, 0.0, 1.0], vertex_colors=[[255, 0, 0], [0, 255, 0], [0, 0, 255]], ), ) # Only update its vertices' positions each frame factors = np.abs(np.sin(np.arange(1, 300, dtype=np.float32) * 0.04)) for i, factor in enumerate(factors): rr.set_time_sequence("frame", i) rr.log("triangle", [rr.components.Position3DBatch(vertex_positions * factor)])

Interpret the same data in several ways using the component level API

The component level API gives you the ability to interpret the same data as multiple types. We do that by logging multiple indicator components, which tell the Rerun Viewer "hey, this entity should be interpreted as type X". In this example we interpret an entity as bott a colored triangle and as three colored points.

import rerun as rr rr.init("rerun_example_manual_indicator", spawn=True) # Specify both a Mesh3D and a Points3D indicator component so that # the data is shown as both a 3D mesh _and_ a point cloud by default. rr.log( "points_and_mesh", [ rr.Points3D.indicator(), rr.Mesh3D.indicator(), rr.components.Position3DBatch([[0.0, 0.0, 0.0], [10.0, 0.0, 0.0], [0.0, 10.0, 0.0]]), rr.components.ColorBatch([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]), rr.components.RadiusBatch([1.0]), ], )

Using your own types with Rerun

The new type oriented API also makes logging data from your own objects more ergonomic. For example, you might have your very own point cloud class.

@dataclass class LabeledPoints: points: np.ndarray labels: List[str])

All you need to do is implement as_component_batches() and you can pass them directly to rr.log. The simplest possible way is to use the matching Rerun archetype’s as_component_batches method like below but you can also get as fancy as you like with custom components and archetypes. Check out the guide on using Rerun with custom data for more details.

@dataclass class LabeledPoints: points: np.ndarray labels: List[str]) def as_component_batches(self) -> Iterable[rr.ComponentBatch]: return rr.Points3D(positions=self.points, labels=self.labels).as_component_batches() ... # Somewhere deep in my code classified = my_points_classifier(...) # type: LabeledPoints rr.log("points/classified", classified)

The main takeaway here is that with 0.9 and the new type oriented API, it becomes a lot easier to use Rerun with your own data types.

How it paves the way for the future

Although this release brings a lot of great updates, it's perhaps the future features it paves the way for that are the most exciting.


Getting data from C++ environments into Rerun was the motivating factor behind the move to our own code generation framework. A large amount of production systems in robotics, computer vision and gaming are built in C++ and we're incredibly excited to soon bring Rerun to all those developers.

Building visualizations inline

Rerun started out making the hard case, where you stream data out of multiple processes and visualize it live, easy. The downside so far has been that in simpler cases, like in jupyter notebooks, using Rerun is more convoluted than it should be.

Even when time is not a factor and you have all your data right there and just want to draw it, you currently have to go through the indirection of logging it first.

The new APIs introduced in 0.9, pave the way for a clean way of just drawing data inline without logging. We'll start rolling that out together with the ability to control layout and visualization options from the SDK later in the year once C++ has landed.

Generating your own Rerun SDK extensions

Our new code generation framework is still a bit immature, but it's been a design goal from the start to let users use it to generate their own stand alone extensions to the Rerun SDK. We hope once it's had time to mature, it will be useful to both teams with their own proprietary data formats and for other projects that want to make interfacing with Rerun as easy as possible for their users.

Let us know what you think

We're incredibly excited to hear what you think about these changes. Join us on Github or Discord and let us know how 0.9 works for you and what you'd like to see in the future.

If you're an existing Rerun user and have any questions or need any help migrating to the new APIs, send us a ping on Discord or elsewhere and we'll be happy to get on a call and help you out.

Rerun OSS beta is released

Today we're making the Rerun open source project public. Rerun is now installable as pip install rerun-sdk for Python users and cargo add rerun for Rust users.

Rerun is an SDK for logging data like images, tensors and point clouds, paired with an app that builds visualizations around that data. We built Rerun for computer vision and robotics developers. It makes it easy to debug, explore and understand internal state and data with minimal code. The point is to make it much easier to build computer vision and robotics solutions for the real world.

Rerun is in beta

Rerun is already quite powerful and useful. A couple of great teams have been using it for several months, as both their main internal debugging tool and as a way to show off their systems to customers and investors. Check out our demo video to get some flavour of what's there now.

We anticipate iterating on core api's and adding core functionality for some time to come. For example, we believe that the ability to configure layout and rendering options from the SDK, read out logged data as DataFrames in Python, and to log data from C++ programs, are must-haves before dropping the beta label. That list can be made a lot longer of course and also includes many data types that don't yet have native support. The beta label is also there to indicate that it's still a young project and there are wrinkles that need smoothing out.

Why make it public now?

We've rebuilt the core of Rerun several times now and building in private where we knew all the users by name has definitely made that easier.

The team has put in a lot of hard work over the last months and we're at a point now where it's already a very useful tool, and the kernel of what Rerun will become is in place. We have a pretty clear idea of where we want Rerun to go, but we want to make sure we go there together with the community. Or better yet, for the community to help us see all the things we're currently missing.

Rerun is now open for contributions and we're all eagerly looking forward to hearing your feedback.

Get involved on Github and come say hi or ask a question on Discord.

Computer vision for tennis

Computer vision is revolutionizing the way we solve problems in the real world. At Rerun, we have the opportunity to work with developers who are creating innovative computer vision products. One company we want to highlight is PlayReplay.

Accurate line calling for everyone

PlayReplay is bringing advanced technology, previously only seen at tournaments on professional tennis courts, to local courts. Their software can accurately call balls in or out and provides in-depth statistics like serve speed, shot distributions and rpm on your spin. The system is self-sufficient, requiring no specialized personnel for each game. It consists of eight sensors that can be attached to the net post and is operated through a tablet or phone. The solution has received endorsement from the Swedish Tennis Federation and has even replaced human line calling in youth tournaments in Sweden.

A video describing the PlayReplay product

An introspection tool for the entire company

When we spoke with PlayReplay's CTO, Mattias Hanqvist, he had big plans for his visualization stack. He wanted to create an introspection tool that could be used across the organization to improve alignment between developers, support teams, and end users. The tool should allow developers to not only understand the algorithms they're working on, but also experience the product as an end user would. It should also enable the first line support team to gain a deeper understanding of algorithms and malfunctioning systems. Additionally, Hanqvist wanted to build visualizations such that they could easily be adapted from being used by developers to being included in the end product for actual users.

PlayReplay's 3D court view, built using Rerun.

Today, PlayReplay is a part of the Rerun alpha user program and has shifted their internal visualizations over to Rerun. The main user for now is their development team. Going forward the plan is to extend the usage across additional parts of the team.

If you’re interested in getting early access to Rerun, then join our waitlist.

From the Evolution of Rosbag to the Future of AI Tooling

Thirteen years ago, Willow Garage released ROS (the Robot Operating System) and established one of the standard productivity tools for the entire robotics industry. Back then, as a member of the ROS core team, I worked on the rosbag storage format and many of the tools it powered. This remains one of the most rewarding and impactful pieces of software I have developed. Rosbag, and other formats like it, are still widely used across the robotics industry, serving a crucial role in enabling developers to understand and evaluate their systems.

Today, the most significant ongoing change in robotics is the use of modern AI in production systems. Creating, training, and integrating the models that power these systems places new demands on development workflows and tooling. Still, the core motivations that drove the early evolution of the rosbag format are as present as ever and influence how we at Rerun think about the next generation of productivity tools for real-world AI. To set the stage for some of that future work, I’d like to share the original story of how the rosbag format evolved.

Rosbag A bag of ROS turtles (credit: Midjourney)

Where It All Started

In 2008, I had the great privilege of joining Willow Garage. Willow Garage was a privately financed robotics research lab with the unique vision of accelerating the entire robotics industry. The fundamental problem Willow Garage wanted to solve was that robotics research and development lacked a functional robotics “stack” to build on top of. Not only did this slow down progress since developers were constantly spending time reinventing the wheel, it dramatically hindered collaboration because developers, left to their own devices, rarely build compatible wheels. Before Willow Garage, I had worked on early autonomous vehicles as part of the DARPA Grand Challenge and aerial sensing systems at Northrop Grumman. I was all too familiar with the pain of building complex systems on top of an ad-hoc home-grown foundation of mismatched tools and libraries.

Willow Garage wanted to solve this problem by building a flexible open-source robotics framework (ROS) and, in parallel, manufacturing a capable robot tailored to the needs of robotics researchers (the PR2). The idea was to build and give away identical robots to ten universities alongside several that would be used in-house by a team of researchers at Willow Garage. By co-developing ROS alongside a shared hardware platform, we could bootstrap an environment where we could get fast feedback not just on the core functionality but things that would push the envelope of collaboration and shared development.

PR2 Beta The first collection of PR2s distributed to the Beta Program (credit: Willow Garage)

(If you want even more ROS and Willow Garage history, check out this IEEE spectrum article.)

Running Code Without Robots

One problem common to robotics is that physical systems severely limit the ease of development. Developers are rarely able to run their code on a live robotic system as quickly and frequently as they might otherwise like to.

If you lack access to a robot or need to share one with others in a lab, you may initially think the robot itself is the bottleneck. However, even at Willow Garage, where we were building a sizable fleet, running code directly on the robot was still quite time-consuming. The need for the robot to interact physically with the world makes it far slower than the average edit-compile-run loop of a software engineer. Moreover, the physical world adds significant non-determinism to the system, making it challenging to recreate a specific scenario where something unexpected might have happened.

One way of addressing this is to stop using real robots altogether and switch to faster-than-realtime simulators. Simulators were also a big part of the solution space within ROS and continue to be an area of investment across the robotics industry today. However, the fidelity of simulators is often limited compared to running in a real system, and high-fidelity simulators eventually start to run up against similar resource constraints.

The other way of addressing this is to derive as much value as possible from whatever time is available to run code live on the robot. The initial goal of rosbag was to create this added value by enabling developers to easily record data about the live robot session, move that data somewhere else, and then later seamlessly playback that recorded data to make use of it in different ways.

Basic Record and Playback

On the recording side, the core architecture of ROS greatly simplified the design of rosbag. ROS encourages users to follow a distributed microservices architecture. Systems are split into separable compute elements, called "nodes," which communicate over shared busses called "topics." Any node can publish a message to a topic, and other nodes interested in those messages can receive them.

Ros Network A ROS network and the rosbag recorder

This architecture makes it straightforward to produce a recording of everything meaningful happening in the system by listening passively in on all of the available topics. As part of the IPC layer, the publishing nodes are responsible for serializing the messages into discrete binary payloads. Rosbag can then receive these payloads, associate them with meta information such as timestamp, topic-name, and message-type, and then write them out to disk. This data was all the early versions of rosbag needed to contain: a repeating series of timestamped message payloads in a single file. We called this a ".bag" file.

Rosbag Format v1.1 Overview of the rosbag V1.1 format

In contrast to recording, the playback function does this in reverse. Rosbag sequentially iterates through the saved messages in the bag file, advertises the corresponding topics, waits for the right time, and then publishes the message payload. Because rosbag directly uses the serialized format to store and transmit messages, it can do this without knowing anything about the contents. But for any subscribers in the system, the messages look indistinguishable from those produced by a live system.

This model established a powerful pattern for working with the robots:

  1. Any time a developer ran code on the robot, they would use rosbag to capture a recording of all the topics. Rosbag could write all of the data directly to robot-local storage. At the end of the session, the recording could be transferred off the robot to the developer's workstation, put in longer-term storage, or even shared with others.
  2. Developers could later inspect the data using the same visualization and debug tools as the live system but in a much more user-friendly context. One of the most powerful aspects of this offline playback is that it enabled additional functionality such as repeating, slowing down, pausing, or stepping through time, much as one might control the flow of program execution in a debugger.
  3. This offline playback could further be combined with new and modified code to debug and test changes. Playing back a subset of the bag file, such as the raw sensor data, into new implementations of algorithms and processing models enabled developers to compare the new output with the initial results. Direct control of the input and the ability to run in a workstation-local context made this a superior development experience to running code on a live robot.

Developers could sometimes go for multiple days and hundreds of iterations of code changes while working with the data from a single recording.

Enhancements to a Bag-centric Workflow

As developers started doing a more significant fraction of their development using bags instead of robots, we started running into recurring problems that motivated the next set of features.

Because ROS1 encoded messages using its own serialization format (conceptually similar to google protobuf), it meant doing something with a bag file required your code to have access to a matching message schema. While we had checks to detect the incompatibility, playing back a file into a system with missing or incompatible message definitions would generate errors and otherwise be unusable. To solve this, we started encoding the full text of the message definition itself into the bag during recording. This approach meant generic tools (especially those written in python, which supports runtime-loading of message definitions) could still interpret the entire contents of any bag, regardless of its origin. One tool this enabled was a migration system to keep older bags up-to-date. When changing a message definition, a developer could register a migration rule to convert between different schema iterations.

Additionally, recording every message in the system meant bags could end up containing millions of messages spread over tens or hundreds of gigabytes. But often, a developer was only interested in a small portion of the data within this file when testing their code. Very early versions of the rosbag player allowed a user to specify a start time and duration for playback. Still, even if rosbag skipped the steps related to reading and publishing, the format required sequentially scanning through every message in the file to find where to start. To address this, we began including a time-based index of all the messages at the end of the file. Not only did this allow the reader to jump directly to the correct place to start playing, but it also made it much easier to only play back a subset of the topics.

Rosbag Format v1.2 Overview of the rosbag V1.2 format

During recording, rosbag still wrote records incrementally to the file, just as before. However, when closing the file, rosbag would append a consolidated copy of the message definitions and indexes at the very end. Since the index is just a performance optimization for data already written to the file, the rosbag tool could regenerate it by scanning the file sequentially from the beginning. This structure meant files could still be recovered and re-indexed if a crash occurred mid-recording -- a vital property for recordings that might take several hours to produce.

More Than Just Playback

Making rosbags easier to migrate and faster to work with ultimately meant developers found themselves with more bags. The need to filter, split, crop, merge, curate, and share bags became increasingly commonplace. Additionally, users often wanted a high-level overview of a bag to figure out where to seek during playback. While the existing playback tool could support these operations crudely, we eventually needed a new approach to working with bags more directly.

We developed another revision of the rosbag format to solve this problem and created new libraries to support a more direct access model. Most notably, this version of the format introduced an internal partitioning called a "Chunk." Chunks are blocks of messages, usually grouped by time or message-type and optionally compressed. Each chunk importantly includes a sub-index describing the messages it contains. Rather than a single index, the end of the bag now had a collection of Chunk Info records with each chunk's location and higher-level information about its contents.

Rosbag Format v2.0 Overview of the rosbag V2.0 format (not to be confused with the ROS2 bag format)

This format and the new libraries enabled us to build new GUI tools like rxbag that could directly open and view the bag's contents. These tools included functionality like per-topic overviews of messages over time, thumbnail previews generated by random access to strategic locations, inline plotting of values from specific topics, and fast scrubbing both forward and backward in time. At a minimum, the tool could use the included schema to create a preview of any message in the file with a textual json-like representation. After using these tools to find a location of interest, a developer could still publish the relevant portion of the stream using the traditional playback model.

qrt_bag A Screenshot from the rqt_bag inspection tool.

These lower-level access libraries also opened up new possibilities. For example, a developer could now write more "batch-style" offline processing jobs without the constraints of the runtime system. This style of job was more suited to higher-performance filtering or transformation tasks.

Good Enough for a Generation

And that was when rosbag generally crossed a critical usability threshold. It was good enough for what people needed to do with it. The rosbag format has remained relatively unchanged since mid-2010 as a foundational piece of the ROS ecosystem. Although replaced in ROS2 with alternative storage mechanisms, it is still the recording format for ROS1, which won't see end-of-life until 2025 -- a remarkable 15-year run of powering robotics data collection, inspection, and playback use-cases.

To summarize, there are a few things that made rosbag truly useful for the ROS community:

  • It is nearly free to use (if you are using ROS). By taking advantage of existing messages in the system, rosbag requires no extra code to start making valuable recordings.
  • Similarly, a powerful playback system also makes it easy to use the data in conjunction with existing algorithms while bringing additional functionality related to control of time and repeatability.
  • Each bag exists as a singular self-contained file with no assumptions about the computer, path, or software context that produced it. You can freely move, save, and share it as an atomic unit. The included schema means the data can always be extracted, even working with files from older versions.
  • The built-in index means rosbag tools function efficiently for common use cases, even when dealing with large files.
  • Direct-manipulation libraries allow developers to process .bag files in alternative ways outside the core ROS execution paradigm.
  • More than just a file format, an ecosystem of generic tooling brings convenient support for visualization, inspection, and file manipulation.

Beyond Rosbag

Perhaps the most controversial aspect of the original rosbag format was the intrinsic dependency on the ROS message definition format and the ROS framework. On the one hand, this choice meant all users of ROS could more easily work with any bag file, improving interoperability within the community. On the other hand, this precluded adoption from users unable to take on such significant dependency. An all too familiar story is startups and research labs, again, rolling their own rosbag-like format from scratch, precluding interoperability with standardized tools.

As part of a move to better interoperate with commercial use cases, ROS2 transitioned to a generic middleware architecture supporting alternative transport and message serialization backends. This transition also introduced a storage plugin system for rosbag instead of a single definitive format. The initial storage plugin for rosbag2 leveraged sqlite3. While the sqlite3 storage plugin takes advantage of a mature open-source library and brings more generalized support for indexing, querying, and manipulating files, it also misses out on some of the original rosbag features. A few downsides of this direction are that it gives up streaming append-oriented file-writing and lacks self-contained message schemas.

At this point, the most logical evolutionary descendent of rosbag is the open-source MCAP format, developed by foxglove as both a standalone container format and an alternative storage backend for rosbag2. MCAP directly adopts many of the original rosbag file format concepts while further adding support for extensible message serialization. It also cleans up a few of the rough edges, such as being more explicit about separate data and summary sections, avoiding the need to mutate the rosbag header, and moving to a binary record-metadata encoding.

MCAP Format Overview of the MCAP format

Further removed from ROS, other formats are similarly motivated by the need for a portable, stream-oriented, time-indexed file format. For example, the facebook research VRS (Vision Replay System) format was developed more specifically for AR and VR data-capture use cases but aims to solve many of the same problems. Rather than using existing serialization formats and generalizing the schema support, VRS uses its own data layout encoding. While this adds a lot of power to the VRS library regarding data versioning and performance, it also adds a fair bit of complexity. Similar to the ROS2 approach with sqlite3, VRS abstracts this complexity behind an optimized cross-platform library. However, this approach makes it challenging to work with a raw VRS file without using the provided library to process it.

What's Needed for the Future?

Great log-replay tools have become an essential part of building successful robotics and perception systems for the real world. Formats like MCAP and VRS give us a great foundation to build upon but are unlikely to be the end of this evolutionary arc.

There have been numerous developments in the years following the design of the rosbag format. An explosion of sensors and energy-efficient computing has unlocked new possibilities for hardware availability, while AI has gone from promising to breathtakingly powerful. The opportunities to solve real-world problems with these technologies have never been greater. At the same time, teams need to be more efficient with their resources and get to market faster.

So what features do the next generation of tools for real-world I need to incorporate?

Capture More than Runtime Messages

The current generations of log-replay tools build on top of schematized “message-oriented” systems like ROS. While this paradigm makes capturing data along the interface boundaries easy and efficient, it can introduce significant friction when capturing data from deeper within components or outside the system runtime. Future tools must give developers enhanced visibility across contexts ranging from core library internals to prototype evaluation scripts.

Easily Express Relationships Between Data Entities

When analytics and visualization tools know about the causal, geometric, and other relationships between data entities, they can provide developers with powerful introspection capabilities. Making it easy for developers to express these relationships effectively requires making them first-class citizens of the data model in a way that is difficult to achieve with current log-replay tools.

Support More General and Powerful Queries

The time-sequential index of events in a system only tells part of the story. To fully understand and explore the relevant data, developers need access to sophisticated and interactive querying and filtering operations. The next generation of tools should allow users to compare and aggregate across runs or any other dimension easily.

Robotics and AI/ML Workflows are Merging

Established practices from the robotics industry are mixing with workflows from the AI/ML community. This will continue and is necessary for achieving widespread success with AI in the real world. Both traditions will need to learn from the strengths of the other.

At Rerun, we’re building a new generation of tools informed by these old lessons and new demands. Much like in the early days of ROS, it’s clear there’s an opportunity to transform the workflows of a new generation of teams. As with the development of rosbag, the steps we take along the way will need to be iterative and incremental, driven by real feedback and use.

In order to work as closely as possible with the community, we plan to open source Rerun in the near future. If you're excited about being a part of that journey, sign up for our waitlist to get early access or a ping when it’s publicly available.

Computer vision for the blind

Computer vision is a powerful technology solving real problems in the real world, already today. It holds the potential to significantly improve life on earth over the next decades. At Rerun we have the privilege to work directly with developers that are building that future. From time to time we will introduce companies building computer vision products for the real world. The first company we want to introduce is biped.

Translating vision to audio

biped is a Swiss robotics startup that uses self-driving technology to help blind and visually impaired people walk safely. The shoulder harness they develop embeds ultra wide depth and infrared cameras, a battery and a computation unit. The software running on the device estimates the positions and trajectories of all surrounding obstacles, including poles, pedestrians or vehicles, to predict potential collisions. Users are warned with 3D audio feedback, basically short sounds similar to the parking aid in a car, that convey the direction, elevation, velocity and type of obstacle. The device also provides GPS instructions for a full navigation assistance.

biped is one of the most advanced navigation devices for pedestrians, and seeks to improve the independence of blind and visually impaired people across the world.

A demo of biped being used out in the wild, including audio feedback users receive

A complex pipeline with complex data

On a high level, biped’s software does the following in real-time:

  1. Sequentially acquire image and depth data from all cameras, fusing the different inputs into a single unified 3D representation
  2. Run perception algorithms such as obstacle segmentation or object detection
  3. Prioritize the most important elements based on risks of collision
  4. Create 3D audio feedback to describe the prioritized elements of the environment

A small change anywhere in the pipeline can affect the performance of downstream tasks and thus the overall performance significantly. For example, the quality of environment understanding strongly affects the prioritization algorithm. biped employs different strategies to counter this problem and to make development as easy as possible. One of them is to visualize intermediate steps of the pipeline with Rerun. This allows the development team to get a quick understanding of how each change affects the whole pipeline.

A visualization of biped’s perception algorithms, built using Rerun.

biped is a part of the Rerun alpha user program and has recently shifted their internal visualizations to Rerun.

If you’re interested in getting early access to Rerun, then join our waitlist.

Inspired by Bret Victor

In his 2014 talk Seeing Spaces, Bret Victor envisioned an environment where technology becomes transparent, where you effortlessly see inside the minds of robots as you build them. This is the dream of everyone building computer vision for the physical world, and is at the core of what we're building at Rerun.

Robot on a table with data visualization on table and walls

A depiction of what a Seeing Space might look like from Bret Victor’s talk Seeing Spaces

Like most interesting people Bret is hard to summarize, but you might say he’s a designer/engineer turned visionary/researcher that talks a lot about interfaces and tools for understanding. He's on a life-long mission to change how we think and communicate using computers.

You know a body of work is special when just taking a small aspect of it, potentially out of context, still produces great ideas. This is without a doubt true of Bret Victor's work, which has been the inspiration for Figma, Webflow, Our World in Data, and many others.

The overlooked inspiration behind Rerun

The best articulation I know of the need to see inside your systems, particularly those with a lot of internal complexity, comes from the talk Seeing Spaces. It’s seldom referenced but I keep going back to it and am always struck by how prescient it was back in 2014.

The full talk Seeing Spaces by Bret Victor.

The context of the talk is roughly the future of maker spaces. In it he makes two main points:

  1. For a growing number of projects with embedded intelligence (robotics, drones, etc), the main challenge isn’t putting them together, but understanding what they are doing and why. What you need here are seeing-tools and we don't really have many of those.
  2. If you’re really serious about seeing, you build a dedicated room (think NASA’s mission control room). We therefore need Seeing Spaces. These spaces would be shared rooms that embed all the seeing-tools you need, similar to how maker spaces have a shared set of manufacturing equipment.

Overall view of the Shuttle (White) Flight Control Room (WFCR) in Johnson Space Center’s Mission Control Center (MCC)

NASA's Shuttle Control Room is built for serious seeing. Photo Credit: NASA

A physical space for seeing is interesting, and if you follow Bret’s work you can see the lineage from this, through The Humane Representation of Thought, to his current project Dynamicland. Whether or not creating a dedicated physical space is the right way to go, for most teams it’s not practical or the top priority. The first problem is getting “regular” software seeing-tools in place that make it easier to build and debug intelligent systems.

This is essentially what we are doing at Rerun. We are building software based seeing-tools for computer vision and robotics. For teams that want to go all the way to Seeing Spaces, the building blocks they need will all be there.

What is a seeing-tool?

Seeing-tools help you see inside your systems, across time and across possibilities. Seeing inside your systems consists of extracting all relevant data, like sensor readings or internal algorithm state, transmitting it to the tool, and visualizing it. This should all be built in and require no additional effort.

Robot visualized on a wall with alternative traveled paths

A depiction of what seeing across time might look like from Bret Victor’s talk Seeing Spaces

Seeing across time means visualizing whole sequences, and making it possible to explore them by controlling time. These sequences could either take place in real world time, or in compute time like steps in an optimization. Seeing across possibilities means comparing sequences to each other, for example over different parameter settings. When training machine learning models, this is usually called experiment tracking.

In essence, a seeing tool is an environment that lets you move smoothly from live interactive data visualization to organizing and tracking experiments.

Principles for a computer vision focused seeing-tool

Every team that builds computer vision for the physical world needs tools to visualize their data and algorithms, and currently most teams build custom tools in-house. Prior to Rerun, we've built such tools for robotics, autonomous driving, 3D-scanning, and augmented reality. We believe there are a couple of key principles we need to follow in order to build a true seeing-tool that can unlock progress for all of computer vision.

Separate visualization code from algorithm code

It's tempting to write ad-hoc visualization code inline with your algorithm code. It requires no up-front investment; just use OpenCV to paint a picture, and show it with cv.imshow. However, this is a mistake because it makes your codebase hard to work with, and constrains what and where you can visualize.

If you instead keep your visualizations separate, it both keeps your codebase clean and opens up for more powerful analysis. It works for devices without screens and you can explore your systems holistically across time and different settings.

You can't predict all visualization needs up-front

For computer vision, visualization is deeply intertwined with understanding. As developers build new things, they will invariably need to visualize what they are doing in unforeseen ways. This means it needs to be easy to add new types of visualizations without having to modify the visualizer or the supporting data infrastructure. This means we need powerful and flexible primitives and easy ways to extend the tools.

When prototyping, a developer should for instance be able to extend a point cloud visualization with motion vectors without recompiling schemas or leaving their jupyter notebook.

The same visualizations from prototyping to production

Algorithms tend to run in very different environments as they progress through prototyping to production. The first prototype code might be written in a Colab notebook while the production environment could be an embedded device on an underwater robot. Giving access to the same visualizations across these environments makes it easier to compare results and removes duplicated efforts.

The increased iteration speed this has can be profound. I've personally experienced the time needed to go from observed problem in production, to diagnosing and designing a solution, and finally deploying a fix, decreasing from days down to minutes.

Why do seeing-tools matter?

Seeing-tools are needed to effectively understand what we are building. They enable our work to span from tinkering to doing experimental science. It’s currently way too hard to build great computer vision based products for the physical world, largely due to the lack of these tools.

The recent progress in AI has increased the amount of people that work on AI powered products. As any practitioner in the field knows, the process of building these products is less like classic engineering, and more a mix of tinkering and experimental science. As we as a community start deploying a lot more computer vision and other AI in real world products, great seeing-tools will be what makes products succeed. At Rerun we made it our mission to increase the number of successful computer vision products in the physical world. And to get there we're building seeing-tools.

If you're interested in what we're building at Rerun, then join our waitlist or follow Rerun on Twitter.

Why Rust?

I've been a programmer for 20+ years, and few things excite me as much as Rust. My background is mostly in C++, though I have also worked in Python and Lua, and dabbled in many more languages. I started writing Rust around 2014, and since 2018 I've been writing Rust full time. In my spare time I've developed a popular Rust GUI crate, egui.

When I co-founded Rerun earlier this year, the choice of language was obvious.

At Rerun we build visualization tools for computer vision and robotics. For that, we need a language that is fast and easy to parallelize. When running on desktop we want native speed, but we also want to be able to show our visualization on the web, or inline in a Jupyter Notebook or IDE.

By picking Rust, we get speed that rivals C and C++, and we can easily compile to Wasm. By using Rust for both our frontend and backend, we have a unified stack of Rust everywhere, simplifying our hiring.

Speaking of hiring, we hoped that by picking Rust, we would attract more high quality developers. This bet on hiring turned out even better than we had hoped.

Sure, but why, really?

Ok you got me! Those are only a part of the reasons we chose Rust. If I'm honest, the main reason is because I love Rust.

I believe Rust is the most important development in system programming languages since C. What is novel is not any individual feature ("Rust is not a particularly original language"), but the fact that so many amazing features have come together in one mainstream language.

Rust is not a perfect language (scroll down for my complaints!), but it's so much nicer than anything else I've used.

I'm not alone in loving Rust - Rust has been the most loved language in the Stack Overflow Developer Survey for seven years straight. So what are the features that make me love Rust so much?

Safety and speed

"Wait, that's two features!" - well yes, but what is novel is that I get both.

To be clear: what I'm talking about here is memory safety, which mean handling array bounds checks, data races, use-after free, segfaults, uninitialized memory, etc.

We've had fast languages like C and C++, and then we've had safe languages like Lisp, Java, and Python. The safe languages were all slower. Common wisdom said that a programming language could either be fast or safe, but not both. Rust has thoroughly disproved this, with speeds rivaling C even when writing safe Rust.

What's even more impressive is that Rust achieves safety and speed without using a garbage collector. Garbage collectors can be very useful, but they also tend to waste a lot of memory and/or create CPU spikes during GC collection. But more importantly, GC languages are difficult to embed in other environments (e.g. compile to Wasm - more on that later).

The big innovation leading to this "fast safety" is the borrow checker.

The borrow checker

The Rust Borrow Checker has it's roots in the Cyclone research language, and is arguably the most important innovation in system program languages since C.

The gist of it is: each piece of data has exactly one owner. You can either share the data or mutate it, but never both at the same time. That is, you can either have one single mutating reference to it, OR many non-mutating references to the data.

This is a great way to structure your program, as it prevents many common bugs (not just memory safety ones). The magic thing is that Rust enforces this at compile-time.

A lot of people who are new to Rust struggle with the borrow checker, as it forbids you from doing things you are used to doing in other languages. The seasoned Rustacean knows to cut along the grain, to not fight the borrow checker, but to listen to its wisdom. When you structure your code so that each piece of data has one clear owner, and mutation is always exclusive, your program will become more clear and easy to reason about, and you will discover you have fewer bugs. It also makes it a lot easier to multi-thread your program.


Rust's enums and exhaustive match statement are just amazing, and now that I'm using them daily I can barely imagine how I could live without them for so long.

Consider you are writing a simple GUI that needs to handle events. An event is either a keyboard press, some pasted text, or a mouse button press:

enum Event { KeyPress(char), Pasted(String), MouseButtonPress { pos: Pos2, button: MouseButton, } } fn handle_event(event: Event) { match event { Event::KeyPress(c) => {} Event::Pasted(text) => {} Event::MouseButtonPress{ pos, button } => {} } }

If you add another alternative to enum Event, then handle_event will fail to compile until you add a handler for that new alternative.

Implementing the above in C or C++ would be very difficult and error prone (and the very existence of std::variant makes me weep in despair).

Error handling

Error handling is an extremely important aspect of the job of an engineer, and failure to report errors can lead to very serious bugs.

In C and Go you have to manually check and propagate errors:

obj, err := foo() if err != nil { return 0, err } result, err := if err != nil { return 0, err }

This is extremely verbose and it is easy to forget an error.

In languages with exceptions, like C++, Java, and Python, you instead have the problem of invisible errors:

auto result = foo().bar();

As a reader, I can't see where potential errors can occur. Even if I look at the function declaration for foo and bar I won't know whether or not they can throw exceptions, so I don't know whether or not I need a try/catch block around some piece of code.

In Rust, errors are propagated with the ? operator:

let result = foo()?.bar()?;

The ? operator means: if the previous expression resulted in an error, return that error. Failure to add a ? results in a compilation error, so you must propagate (or handle) all errors. This is explicit, yet terse, and I love it.

Not everything is perfect though - how error types are declared and combined is something the ecosystem is still trying to figure out, but for all its flaws I find the Rust approach to error handling to be the best I've ever used.

Scoped resource management

Rust will automatically free memory and close resources when the resource falls out of scope. For instance:

{ let mut file = std::fs::File::open(&path)?; let mut contents = Vec::new(); file.read_to_end(&mut contents)?;// when we reach the end of the scope, // the `file` is automatically closed // and the `contents` automatically freed. }

If you're used to C++ this is nothing new, and it is indeed one of the things I like the most about C++. But Rust improves this by having better move semantics and lifetime tracking.

This feature has been likened to a compile-time garbage collector. This is in contrast with a more common runtime garbage collected language, where memory is freed eventually (at some future GC pass). Such languages tend to use a lot more memory, but worse: if you forget to explicitly close a file or a socket in such a language, it will remain open for far too long which can lead to very subtle bugs.


I find WebAssembly (or Wasm for short) a very exciting technology, and it probably deserves a blog post on its own. In short, I am excited because with Wasm:

  • I can write web apps in another language than JavaScript
  • I can write web apps that are fast
  • I can safely and efficiently sandbox other peoples' code

So what does Wasm have to do with Rust? Well, it is dead easy to compile Rust to Wasm - just pass --target wasm32-unknown-unknown to cargo, and you are done!

And then there is wasmtime, a high performance runtime for Wasm, written in Rust. This means we can have fast plugins, written in Rust, compiled to Wasm, running in our tool. Rust everywhere!


The Rust trait is really nifty as it is the interface for both run-time polymorphism and compile-time polymorphism. For instance:

trait Foo { fn do_stuff(&self); } // Run-time polymorphism (dynamic dispatch). // Here `Foo` acts like an Java interface or a abstract base class. fn runtime(obj: &dyn Foo) { obj.do_stuff(); } // Compile-time polymorphism (generics). // Here `Foo` acts as a constraint on what types can be passed to the function // (what C++ calls a "concept"). fn compile_time<T: Foo>(obj: &T) { obj.do_stuff(); }


Rust has amazing tooling, which makes learning and using Rust a much more pleasant experience compared to most other languages.

First of all: the error messages from the compiler are superb. They point out what your mistake was, why it was a mistake, and then often point you in the right direction. The Rust compiler errors are perhaps the best error messages of any software anywhere (which is fortunate, since learning Rust can be difficult).

Then there is Cargo, the Rust package manager and build system. Having a package manager and a build system for a language may seem like a low bar, but when you come from C++, it is amazing. You can build almost any Rust library with a simple cargo build, and test it with cargo test.

Rust libraries are known as crates (and can be browsed at Though the ecosystem is nascent, there is already a plethora of high quality crates, and trying out a crate is as easy as cargo add. There is of course some legitimate worry that the Rust crate ecosystem could devolve into the crazy left-pad world of npm, and it is something to be wary about, but so far the Rust crates keep an overall high quality.

And then there is the wonderful rust analyzer which provides completion, go-to-definition, and refactoring to my editor.

Rust documentation is also really good, partially because of the effort of its writers, partially because of the amazing tooling. cargo doc is a godsend, as are doc-tests:

/// Adds two numbers together. /// /// ## Example: /// ``` /// assert_eq!(add(1, 2), 3); /// assert_eq!(add(10, -10), 0); /// ``` fn add(a: i32, b: i32) -> i32 { a + b }

The compiler will actually run the example code to check that it is correct! Amazeballs!

The bad

It's not all unicorns and lollipops. Rust has some pretty rough edges, and may not be for everyone.

It's not a simple language

Rust is difficult, and it takes a while to learn. Even if you know C and some functional programming, you still need to learn about the borrow checker and lifetime annotations. Still, I would put Rust as both simpler and easier than C++.

Compile times

This is unfortunately something Rust has inherited from C++. Things are bad, and are only slowly getting better, and I doubt it will ever be fast as e.g. Go.

Noisy syntax

You will see a lot of <'_> and ::<T> in Rust, and it ain't always pretty (but you get used to it).

Floating point behavior

f32 and f64 does not implement Ord. This means you cannot sort on a float without jumping through a lot of hoops, and this is very annoying. I wish the float would just use total ordering and take the performance hit.

Same with Hash, which f32 and f64 also doesn't implement.

Thankfully there is the ordered-float crate, but the ergonomics of using a wrapped type isn't great.

Still lacking a lot of libraries

The Rust crate ecosystem is good, but C and C++ has a huge head start and it will take a long time for Rust to catch up. For us at Rerun, that pain is most urgently felt in the lack of libraries for scientific computing and computer vision, as well as the lack of mature GUI libraries.

Flawed, but improving

Five years ago my gripes with Rust were much longer. Rust is steadily improving, with a new release every six weeks. This is an impressive pace, especially since there are no breaking changes.


At the end of the day, a programming language is a tool like any other, and you need to pick the right tool for the job. But sometimes, the right tool for the job is actually the tool you love the most. Perhaps that is exactly why you love it so much?

In many ways, using C++ for the engine, Go for the backend, and JS for the frontend would have been the "safe" choice. We could have made use of the many great C++ libraries for linear algebra and computer vision, and we could have used one of the many popular and excellent frontend libraries for JS. In the short term that might have been the right choice, but it would have severely limited what we could accomplish going forward. It would have been a bet on the past. At Rerun, we are building the tools of the future, and for that we need to be using the language of the future.

If you're interested in what we're building at Rerun, then join our waitlist or follow me on Twitter.

We’re hiring!

Hi! My name is Emil, and I’m a developer with a passion for Rust, dev tools, user interfaces, and climbing. I’m also the creator and maintainer of a couple of open source projects, including egui.

A few months ago, two friends and I co-founded Rerun. We are building a brand new kind of visualization tool for companies working in computer vision. I am extremely excited by this - not only is it a fun thing to build, I also believe it will have a huge, positive impact on the world.

We’ve raised some money, and now we are looking to hire a founding team of great developers to build the first version of our product.

Rerun’s mission

I met my co-founders Niko and Moritz at a computer vision and 3D scanning company where we wrote some really cool in-house visualization tools. We used these tools to quickly iterate when prototyping new algorithms, investigating weird corner cases, and to troubleshoot our deployed scanners. These tools were critical to our success, but took a lot of effort to write. We’ve since learned that all successful computer vision companies build their own in-house visualization tools. This costs them time and effort, and produces tools that aren’t as good as they should be. We want to change that!

20 years ago, if you had an idea for a game, you had to first write a game engine. This stopped most people, and slowed down the rest. Today there are reusable game engines such as Unity and Unreal, and there has been an explosion of great games coming from thousands of studios and indie developers. At Rerun, we are going to write a reusable visualization toolbox for the entire computer vision industry.

Computer vision is going to diagnose illnesses, enable blind people to read street signs, replace insecticides with bug-zapping lasers, power autonomous drones that do repairs in remote and dangerous areas, and on and on. Computer vision is on the cusp of revolutionizing the world, and Rerun will be at the heart of it.

What we are building

A developer starts by using our logging SDK (for C++, Python, Rust, …) to log rich data as easy as they would log text. Rich data includes images, meshes, point clouds, annotations, audio, time series, and so on.

The logs are transformed and cached by a server which sits between the logging SDK and our viewer.

The viewer is an app that can run both natively and in the browser. It can view the logs live, but also recordings. You can scrub time back and forth and compare different runs, and you can customize the viewer to suit your needs.

For the first year we will be working closely with a few select customers, and we will select the customers that are doing the coolest things!

Working at Rerun

We are going to be a small, tight team, working together to build a great product of the highest quality. We believe in being helpful, curious, honest, and never settling for something less than great.

We are very ambitious. This will be difficult, so you need to be good at your job.

We are looking to hire experienced developers who can make their own decisions while also being part of a collaborative team. We want to create a workplace where you can do your life’s best work and learn new things, and also have a family and hobbies.

We are a distributed company with a remote-first culture. The founders are based in Stockholm, Sweden. We expect everyone on the team to be available for discussions and meetings every weekday 13-17 CET. We plan to get the entire team together for a full week once a quarter.

We will pay you a competitive salary with six weeks of paid vacation. You will also be offered an options/equity package. We will pay for whatever hardware and software you need to do your job. You can of course use the OS and editor you feel most comfortable in.

What are the skills we are looking for?

We are creating tools that are extremely easy to use, that look beautiful, and that run butter smooth. We want Rerun to become the kind of tool that you would switch jobs just to be able to use.

We are looking for people who enjoy being helpful to colleagues, customers, and to the open source community.

We are an open core company, so most of what you do the first few years will be open source. You will be using a lot of open source libraries and should be able and willing to contribute to them via issues and PR:s. Any open source experience is a plus.

We expect you to write clean, readable code, intuitive API:s, and good documentation. Communication skills are essential!

We are building everything in Rust, so you should either know some Rust or have the desire and ability to learn it. If you already know C++ or C, you should be fine. Why Rust? Rust is modern and fast, but most importantly: it runs anywhere. We can have the same code run on desktop, the cloud, on edge devices and in the browser.

It’s great if you have a sense of how images and audio are represented by a computer and a good grasp on linear algebra and statistics.

Any experience building dev-tools is a big plus.

We’re building tools for the computer vision industry, including robotics, drones and self-driving cars. Any experience in relevant fields is a plus, but not a requirement. Knowledge of ROS is also a plus.

We are also very interested in different perspectives, so if you have a different background or experiences than the founding team, you should also apply!

Finally, we want those that have a desire to make a deep, positive impact on the world.


There are a few roles we need to fill, with a lot of overlap between them. You do not need to fit snugly into a role to apply! You also don’t need to tick all of the boxes - you will perhaps bring other talent to the table which we haven’t foreseen the need for!

We are building a UI that is intuitive, beautiful, and responsive. Any experience building editors (video, audio, CAD, …) is a big plus. The UI will be written in Rust. It will be compiled to WASM and run in a browser, so knowledge of web programming is also a plus.

The viewer needs to be able to scroll through and filter big datasets in real-time, so we are looking for people who know how to write fast code. You should have a good sense of how a CPU works. It is a big plus if you have built game engines or other high-performance real-time apps, as is any experience of threading or SIMD.

We need a graphics engineer who can build a renderer that can run in a browser and also leverage the full performance of the GPU on the desktop. This means the graphics must not only look good, it must also scale. We will likely write the renderer on top of wgpu.

We are writing a high-performance server that needs to index and cache large amounts of visualization data. Experience with databases is a plus here, as is knowledge of tokio or other async code.

We are building a Python SDK for logging visualization data. For this, we want someone with experience building quality Python libraries, preferably someone with experience with OpenCV, PyTorch, Tensorflow, and other Python libraries that we are likely to integrate with.

We are also building a C++ SDK, and here we want someone that feels comfortable with building a C++ library with all that entails. Our C++ SDK needs to be able to run on edge devices too, so experience with embedded software is a plus. Both the Python and C++ SDK:s will be interfacing with Rust over FFI.

Join us!

You can read more about or roles on our jobs page. Even if no single role fits you perfectly, don’t worry - just apply for the role that is closest!

We’re looking forward to hearing from you! ❤️

edited on 2022-07-05 to reflect that we now accept remote candidates

Starting Rerun

My career-long obsession with computer vision started with a long distance relationship back as an undergraduate. Over countless hours on Skype, we missed looking each other in the eyes. Because the camera is outside the screen, you always look just past each other. I figured it should be solvable with some math and code, so I decided to learn computer vision to fix it - turns out that was much harder than I thought. I’ve been hooked on solving problems with computer vision ever since.

More than a decade later I’ve started Rerun to build the core tools that are missing for computer vision teams to make their products successful in the physical world. My co-founders are two of the most amazing people I know, both old colleagues and friends, Moritz Schiebold and Emil Ernerfeldt. At Rerun we’re building visualization infrastructure for computer vision in the physical world. AI and computer vision has an incredible potential to improve life on earth and beyond. We believe the deployment of this technology needs to happen much much faster.

Three smiling men posing for a portrait

In 2012 as I was taking Stanford’s Introduction to Computer Vision class with Fei-Fei Li, I had mixed emotions. My mind was opening to the incredible possibilities that computer vision could unlock if it worked, but I also found that in practice it mostly didn’t.

After graduation I joined a 3D body-scanning startup called Volumental, right at the top of the 3D hype cycle. Most of that batch of 3D computer vision startups folded or got acqui-hired, but Volumental persevered. Moritz was the CEO back then and ran a tight ship. We made it through sheer determination to stay alive, an unreasonably strong tech team, and critically to this story, fantastic tooling.

In the early days, we were hand coding each visualization. Doing so was costly in both effort and code complexity so we did as little as we could get away with. What we didn’t know yet was how much everything we didn’t see slowed us down.

When Emil joined us from the gaming industry, he brought a critical–and very different– perspective and skillset. He would start out building an interactive app that visualized everything before touching the actual problem. Visualization Driven Development if you will. Because most of us weren’t strong graphics programmers, Emil built a framework that made it dead simple to build these apps by just providing the data. The increased iteration speed this gave us was key to shipping a product that actually worked, years before any viable competitor, and clearly winning that market.

Think about the map a self-driving car is building: it shows you where the car thinks it is relative to all other objects. You can just look at that and immediately understand how the car’s software is reasoning. You can also quickly understand where it might be wrong. While engineers need more detail, customers and the broader team also need this understanding.

A common misconception is that the rise of deep learning in computer vision makes visualization less important, since models are trained end-to-end and are effectively black boxes. In fact, the opposite is true. Powerful deep learning models make vastly more problems feasible to solve. For all but the simplest applications in the physical world, these models are just components of larger systems. More powerful models also enable more complex behavior that those building the products need to understand. Visualization for computer vision is becoming significantly more important.

Companies like, Weights & Biases, and Hugging Face have made deep learning easier by tackling dataset labeling, experiment tracking, and using pre-trained models. Unfortunately, the toolset for rapid data capture and visualization that was so critical for Volumental, still isn’t broadly available. Every computer vision company building for the physical world currently needs to build it in-house. Many make due with ineffective tools because of the high cost of development. We’re fixing that.

Product screenshot

Over the past decade computer vision research has progressed faster than anyone could have imagined. At the same time, real world problems haven’t been solved at nearly the same rate. The expected revolutions in fields like autonomous vehicles, agriculture, manufacturing, augmented reality, and medical imaging are all behind schedule. With the right tools for iterating on these products quickly, we can speed up that deployment by orders of magnitude. More projects will get off the ground and big projects will go faster.

Getting this right is a huge undertaking and to get started we’ve just raised a $3.2M pre-seed round from Costanoa Ventures, Seedcamp, and amazing angels from all over computer vision. I couldn’t be more excited for this opportunity to help our field truly improve life on earth over the decade ahead!