What it does Rerun is a Rust-built data layer for physical AI: log images, point clouds, joint states, video, and time series from robots, simulators, or CV pipelines. A built-in viewer plays everything back in temporal lockstep, so you can scrub through episodes and compare sensors side-by-side. The same storage is queryable via dataframes or SQL, and streams directly into model training without export jobs.

The interesting bit The storage engine is column-chunked and purpose-built for multi-rate physical data — different sensors logging at different frequencies, all kept in sync. That same substrate powers visualization, querying, and training pipelines, so you’re not maintaining three stale copies of the same robot run.

Key highlights

• SDKs in Python, Rust, and C++; pip install rerun-sdk gets you logging in under two minutes
• Ingests MCAP, .rrd, LeRobot, and other robot log formats natively
• Viewer runs in-browser or as a native app with time-aware scrubbing
• Query APIs to extract clean training datasets from recordings
• Open-core model: all code in the repo stays MIT/Apache; commercial “Rerun Hub” is separate

Caveats

• API is still evolving; the README warns to “expect breaking changes”
• Viewer performance degrades with too many entities or multi-million point clouds (tracked issues #7115 and #1136)
• C++ and Rust SDKs require a separate viewer binary install; only Python bundles it

Verdict Grab it if you’re debugging robot perception, building CV pipelines, or tired of stitching together ROS bags, Jupyter notebooks, and training dataloaders. Skip if you just need a lightweight RViz replacement and don’t care about querying or dataset extraction.