Description
Rerun offers a comprehensive data layer designed for the complexities of physical AI and robotics. Its core purpose is to provide the essential data primitives needed to construct, comprehend, and enhance your data loop. This platform is engineered to manage multi-rate and multimodal data, scaling seamlessly from the very first recording to petabyte-level datasets.
Rerun's capabilities span the entire data pipeline. It excels at visualizing everything, allowing users to review datasets, debug intricate issues, and extend functionality with custom views and tools. For data analysis, Rerun supports full dataframe or SQL queries across any robotics data, enabling deep dives into recordings. The platform also facilitates data transformation without the need for copies, allowing for schema evolution and the addition of derived columns without compromising historical data integrity.
A key differentiator is Rerun's support for training AI models directly on recordings. Users can express dataset mixes as queries and stream data to GPUs, leveraging column-aware and video-codec-aware dataloaders. This eliminates the need for a separate export step, streamlining the machine learning workflow. Furthermore, Rerun fosters collaboration by ensuring everyone works from the same data, with a single viewer and shared recordings accessible across teams.
The Rerun Hub infrastructure powers this data loop, acting as a production backend for cataloging, indexing, and retrieving data from object stores. This transforms raw storage into a queryable, streamable foundation, enabling edge or near-data transformations. The platform is open-source, with the Rerun team actively contributing to related open-source projects like egui and gfx-rs.
Rerun is particularly valuable for researchers and engineers in robotics, computer vision, and AI development. It provides a unified solution for visualization, querying, transformation, and training, significantly speeding up development cycles and improving the debugging process for complex physical AI systems. Its ease of integration across Python, C++, and Rust makes it a versatile tool for a wide range of projects.
Rerun's Core Features
Data primitives for building and understanding data loops
Support for multi-rate and multimodal data
Scalable from initial recording to massive datasets
Comprehensive visualization of datasets
Debugging tools for detailed issue analysis
Extensible with custom views and tools
Full dataframe and SQL querying capabilities
Data transformation without data copies
Schema evolution support without breaking history
Direct model training on recordings
Column-aware and video-codec-aware dataloaders
Shared data access for team collaboration
Open-source with active community contributions
Python, C++, and Rust SDKs available
How to use Rerun?
Get Started: Follow quick start guides for C++, Python, or Rust.
Log Data: Use the Rerun logging SDK to record multi-rate, multimodal data.
Visualize: Open the Rerun Viewer to explore, debug, and annotate your data.
Query Data: Utilize SQL or dataframe queries to analyze recordings.
Transform Data: Refine your data and evolve schemas without creating copies.
Train Models: Stream data directly from recordings to your GPUs for training.
Share Insights: Collaborate with your team using shared recordings and the Rerun Viewer.
Rerun's Use Cases
- Robotics Data Visualization
- AI Model Training
- Data Loop Improvement
- Multimodal Data Analysis
- Real-time Debugging
- Collaborative Development
- Computer Vision Workflows
- Physical AI Research
