// the find
rerun-io/rerun
Visualize, query, and stream to train on multimodal robotics data.
Rerun is a data platform for physical AI and robotics: log multimodal sensor streams (images, point clouds, time series, joint states), visualize them in a purpose-built viewer, and query recordings via SQL or the dataframe API. It targets robotics engineers and CV researchers who need something between a debugger and a data warehouse for multi-rate, heterogeneous sensor data. Think RViz, but with persistent storage, replay, SQL queries, and a training pipeline integration story.
- The columnar chunk storage model genuinely handles multi-rate data correctly — a camera at 30Hz and a lidar at 10Hz are stored and queried without forcing them onto a common timeline, which is the thing every naive robotics logger gets wrong.
- DataFusion integration means you can run SQL over a recorded session without exporting anything. Querying 'show me all frames where detected_objects > 3 and confidence < 0.7' against a real recording is exactly the kind of thing that normally requires a full data pipeline.
- MCAP support means it slots into existing ROS2 tooling without a migration — you can point it at files you already have, which is a practical onramp for teams with years of existing recordings.
- The code generation pipeline (FlatBuffers → Rust + Python + C++ SDKs) keeps the three language SDKs structurally identical. Cross-language interop issues are solved at the schema level, not at runtime.
- The viewer chokes on large entity counts and multi-million-point clouds — they document both as known open issues. For serious robotics workloads (dense lidar, many tracked objects), you will hit these ceilings quickly.
- Breaking API changes are explicitly promised. Adopting this in a production data pipeline today means you're signing up for periodic migration work with no stability guarantees.
- C++ and Rust users need a separately installed `rerun-cli` binary to get the viewer; only Python bundles it. In CI or embedded environments, this is an extra dependency that will bite someone during a late-night build.
- The `nasm` feature for video decoding silently falls back to worse performance if the system CLI isn't installed — no warning at startup, no error, just slower video. That kind of invisible degradation is hard to diagnose in production.