Rerun is an open-source data layer for physical AI: a system to log, query, visualize, and stream multimodal data for building robotics and computer-vision applications. Released under Apache-2.0 and built in Rust, it has passed 11,000 GitHub stars by tackling a problem that quietly slows almost every embodied-AI team: the data that agents and robots produce is multi-rate and multimodal, and the usual patchwork of loggers, viewers, and export scripts turns understanding it into a chore. Rerun replaces that patchwork with one substrate that ingests everything, renders it in sync, and streams it straight into training. ## What It Does Rerun ingests multi-rate, multimodal data, including images, point clouds, transforms, time series, tensors, joint states, and video, from many sources and formats: robot logs, human-data rigs, simulation, and web video, across containers like MCAP, its own rrd files, and LeRobot datasets. Its built-in viewer renders all of it in sync and in real time, so you can scrub through an episode, compare sensors side by side, or watch a computer-vision pipeline run live. Crucially, the same data is queryable with dataframes or SQL and streams directly into training, without a separate export step producing stale copies. ## The Architecture Under the hood, Rerun is built on column-chunk storage purpose-built for multi-rate physical data, implemented in Rust for performance. Data is organized around entities on timelines, so heterogeneous streams that arrive at different rates stay aligned and can be replayed, queried, and visualized coherently. This columnar foundation is what lets the same recording serve three roles at once: a live debugging view, a queryable dataset, and a training input. SDKs are available for Python, Rust, and C++, and a developer can install the Python package with a single pip command and see their first data in the viewer in about two minutes. ## Built for Real Workflows The use cases the project targets are concrete. Teams ingest robot logs, egocentric and UMI rigs, simulation, and web video into a single substrate; run computer-vision pipelines such as SLAM, hand tracking, and motion retargeting as table edits; query raw, intermediate, and derived data with dataframes or SQL; visualize multi-rate multimodal sequences across an entire pipeline; and stream dataset mixes directly to training without export jobs. The through-line is removing the friction between recording data, understanding it, and training on it. ## Usability in Practice Rerun is genuinely quick to start with, especially in Python where the SDK ships bundled with the viewer, while C++ and Rust rely on installing the viewer binary separately. Instrumenting existing code is largely a matter of logging entities to named paths on a timeline, which fits naturally into scripts and notebooks. The honest caveats are that the project is in active development with an evolving API and explicit warnings about breaking changes, and that the viewer can slow down when a scene contains a very large number of entities or multi-million-point clouds. For its target audience of robotics and CV engineers, those trade-offs are usually acceptable given how much manual tooling it replaces. The project also ships agent skills that help coding assistants write Rerun code. ## Pros and Cons The strengths are clear: Rerun unifies ingestion, real-time synced visualization, SQL and dataframe querying, and streaming to training in one tool; its Rust columnar core is fast; it supports Python, Rust, and C++; it reads common robotics formats like MCAP and LeRobot; and the Apache-2.0 license is permissive. The trade-offs are that it is still pre-1.0 with an unstable API and expected breaking changes, the viewer has known performance limits on very large scenes, and the entity-and-timeline data model has a learning curve. It is also purpose-built for multimodal and physical-AI workflows rather than being a general-purpose dashboard. ## Outlook Rerun is riding the surge of interest in physical and embodied AI, where robotics, simulation, and large multimodal datasets are converging and the tooling for handling that data has lagged the models. A single, fast, queryable layer that spans logging, visualization, and training is a natural fit for that moment, and the project's adoption across robotics, simulation, and computer vision suggests it has found real traction. As the API stabilizes toward a 1.0, its appeal to production teams should broaden further. ## Conclusion Rerun is a well-conceived, high-performance open-source data layer for teams working with multimodal, physical-AI data. It is best suited to robotics, simulation, and computer-vision engineers who need to record, understand, query, and train on complex sensor streams, and less relevant to those seeking a conventional business-metrics dashboard. For anyone building at the intersection of multimodal data and physical AI in 2026, it is one of the most capable open tools available today.