Rerun

Rerun란 무엇인가요?

Rerun is a data layer for Physical AI and robotics workflows. It is built for multi-rate, multimodal data, with tools to log, store, query, visualize, transform, and train from the same underlying data model.

The open-source SDK covers local development and analysis, while Rerun Hub adds a managed catalog and storage engine for teams that need persistent, single-tenant infrastructure over their own object storage. Together, they support a workflow that starts with a first recording and can scale to queryable datasets and training streams.

Rerun(으)로 무엇을 할 수 있나요?

Single data model across the loop

Log, store, query, transform, visualize, and train on the same multi-rate, multimodal data model instead of moving between separate tools and formats.

Interactive visualization

Inspect synchronized robotics data in the Viewer, including cameras, joint states, and 3D motion, to debug episodes and review datasets.

Query and reshape data

Run SQL or dataframe queries across recordings, including columns, time ranges, and values inside the data itself, not just metadata.

Recording and catalog storage

Store recordings as column-chunks in `.rrd` files and register them into a catalog for queryable segments across one or many recordings.

Training from queries

Stream dataset mixes directly to training jobs with a column-aware, video-codec-aware dataloader that works from catalog queries.

Local or managed deployment

Use the platform either locally with the open-source SDK or with Rerun Hub for managed, single-tenant storage and retrieval in the cloud.

사용 사례

“Robot debugging and episode review”

Record sensor data from robotics systems, then review synchronized views of cameras, joints, and motion in the Viewer to understand what happened in an episode.

“Dataset analysis and transformation”

Query recordings with SQL or dataframe operations to inspect values across time ranges, columns, and segments, then derive new layers such as labels or scores.

“Importing existing data”

Convert data from existing formats such as MCAP, ROS 2, or Parquet into `.rrd`, then keep the whole workflow in one format and one catalog.

“Collaborative review and annotation”

Point a team at the same recordings through a shared viewer and catalog so they can annotate, investigate failures, and trace issues back to source data.

“Training from recorded data”

Build training batches directly from catalog queries so models learn from the same recordings that were logged and inspected, without a separate export format.

자주 묻는 질문

How do you get started with Rerun?

Rerun’s SDK is installed with `pip install rerun-sdk`, and the quick-start docs also show an optional `[dataplatform, dataloader]` install for queries and training features. For Rust and C++, the site points to separate install and project setup docs.

What workflow does Rerun support?

The site says the SDK can log, query, transform, visualize, and train on multi-rate, multimodal data. The quick-start flow is to log or convert data into `.rrd`, register it in a catalog, and then visualize, query, or train from that catalog.

What file formats can Rerun work with?

The web viewer opens `.rrd` recordings directly in the browser, and the viewer page says it is for opening and exploring `.rrd` recordings online. The quick-start docs also note support for `.mcap` files and more.

Is Rerun free to use?

The pricing page separates the open-source SDK from Rerun Hub. The SDK is free and open source on your own machine, while Hub is a managed, single-tenant cloud offering that runs the same data layer at scale in your region.