Rerun is a data layer for Physical AI scenarios such as robotics, autonomous driving, drones, and space AI. It aims to put multi-rate, multimodal data—from capture, logging, visualization, querying, and transformation through to training—onto a single data model. Its core file format is the columnar .rrd, accompanied by a Viewer, SDK, CLI, local catalog, and commercial Hub.
Functionally, the Rerun SDK can log data such as sensors, images, videos, point clouds, scalars, and spatial transforms directly from code, and it can also convert existing formats into .rrd. The Viewer supports both desktop and web, and is used for reviewing datasets, debugging detailed issues, and extending custom views. The query layer supports SQL and DataFrame access, allowing records to be queried by column, time range, and value rather than only by metadata. On the training side, it provides data-loading capabilities for PyTorch DataLoader, enabling batches to be streamed directly from the catalog. In terms of languages, it explicitly supports Python, Rust, and C++, and mentions integrations with ROS 2, MCAP, LeRobot, notebooks, web embedding, and S3-compatible object storage.
The Rerun SDK is an open-source toolchain dual-licensed under Apache-2.0 / MIT. The company says it is free forever, including logging, query, transform, visualize, train, local catalog, and community support. The commercial Rerun Hub provides a persistent hosted catalog, byte-range indexing, cross-object-storage queries, team sharing, authenticated links, SSO, and single-tenant isolation. Pricing is not public and requires contacting sales; contracts are based on deployment, team size, and data scale. For self-hosting, the SDK and local catalog can be run independently. Hub is mainly hosted by Rerun in the customer’s chosen cloud region, while larger teams can discuss deploying the data plane into their own account.
Its main strength is that the free SDK is feature-complete, making it especially suitable for robotics data workflows involving time-synchronized, multimodal, and spatially dense data. The documentation covers Getting Started, APIs, migration, MCAP, querying, training, and integrations, and is generally high quality. The downsides are that the product is vertically positioned, so ordinary web/backend development teams may not need it; Hub pricing is opaque; full self-hosting options are limited; and the training dataloader section is marked experimental in examples, so it needs production validation. It is best suited to teams working on robotic perception, control, planning, ML dataset construction, and Physical AI R&D.
The captured text does not provide information on mainland China network access, payments, or local compliance, so its accessibility status is unknown. If network access or procurement is restricted, alternatives to evaluate include the ROS 2 toolchain, Foxglove, Open3D, Weights & Biases, or a custom data pipeline built on object storage.
⚠ This review is compiled from public sources and does not constitute a purchase recommendation. Verify all facts on the vendor's official site. Verify on rerun.io official site.
rerun.io is an United States Dev Tools provider. TG4G tracks its product information, an overall rating of 8.0/10, and a China-accessibility score of Workable. Click "Visit Official Site" to reach rerun.io directly.