Rerun — The Data Layer for Physical AI

TL;DR

Rerun is an open-source toolchain (SDK + viewer) plus a commercial backend (“Rerun Hub”) positioned as a unified data layer for Physical AI / robotics: logging, querying, visualizing, and training on multi-rate, multimodal data (2D, 3D, text, time series, tensors). The 0.32 SDK release reframes Rerun from a visual-debugging tool into a full data layer with dataframe/SQL access over raw and derived data, a column-chunk storage format, and edge/near-data transforms. Rerun Hub adds a catalog, byte-range indexing, and retrieval over object stores so the same logs are streamable and queryable for training.

Key claims

A single open-source toolchain (pip install rerun-sdk) covers logging, querying, visualization, and training data prep for multi-rate, multimodal robotics streams [landing page §Open source SDKs].
The 0.32 SDK release positions Rerun as a unified data layer rather than just a visualizer — adding dataframe + SQL access across raw, intermediate, and derived data [landing page §News, “A new data layer for robot learning”].
All data is stored in a column-chunk format that the rest of the stack is built on; logging, storage, and streaming share the same primitive [landing page §Log, store, and stream].
Rerun Hub is the commercial backend: catalog + byte-range indexing + retrieval that turns object stores into a queryable, streamable foundation, with the ability to run transforms on the edge or close to the data [landing page §Rerun Hub].

Method

Three layers, per the landing page. (1) SDKs — language bindings that log multi-rate streams (images, point clouds, transforms, scalars, tensors) into a typed column-chunk format, with a visualizer that scrubs along time and lets users build custom views over the pipeline. (2) Query / transform — dataframe and SQL surfaces over raw, intermediate, and derived data, so adding a column or slicing across runs feels like editing a table. (3) Rerun Hub — production backend that indexes the column chunks in object storage, provides byte-range retrieval, and runs transforms either on the edge (close to capture) or close to the catalog. The 0.32 release is the cut where these three layers compose into a single data layer rather than a visualizer with bolt-ons.

Results

No benchmarks; this is a product launch page. Concrete proof points are the 0.32 release notes (linked from the page) describing the unified architecture, and an existing user surface that includes robot-capture tooling like the Robocap repo (see SLAM on Robocap via cuVSLAM, packaged with rattler-build (Claude Code build report) for a real cuVSLAM-on-Robocap build report visualizing into Rerun).

Why it’s interesting

Rerun is the data-infra companion to a lot of the robotics / world-model / capture work the wiki has been accumulating. The “log once in a column-chunk format, then query / visualize / train off the same store” pitch is the inverse of the common pattern where capture pipelines write opaque blobs and training pipelines re-ingest them — and that re-ingest step is where most of the engineering pain sits for embodied / multimodal datasets. Worth tracking as the underlying substrate for several adjacent items on the wiki: it is the visualization target in SLAM on Robocap via cuVSLAM, packaged with rattler-build (Claude Code build report)‘s cuVSLAM-on-Robocap pipeline, the kind of “incremental-layer perception pipeline” host that would naturally sit under Fauna Sprout: A lightweight, approachable, developer-ready humanoid robot-style developer-robot workflows, and a candidate front-end for browser-side embodied demos like mjswan: real-time interactive MuJoCo + ONNX policy simulation in the browser. The Hub piece (catalog + byte-range retrieval over object stores) also rhymes with the data-engineering theme in On Data Engineering for Scaling LLM Terminal Capabilities and Synthetic Data Powering Pretraining (UC Berkeley EE 290/194-11 Lecture 11a) — the bottleneck on training Physical AI is increasingly “how is the data laid out”, not “what is the model”.