mosaico-labs/mosaico

What it solves

Mosaico addresses the difficulty of managing and retrieving data for robotics and Physical AI. It solves the problem of monolithic, unsynchronized sensor logs (like ROS bags) that are difficult to query and incompatible with the dense, synchronous tensor formats required for training modern AI models.

How it works

The platform uses a client-server architecture consisting of a Rust-based backend daemon (mosaicod) and a Python SDK. It transforms raw sensor data into a structured, queryable archive using a zero-copy architecture to enable direct random access to signals without parsing entire files. The system employs a strictly-typed data ontology to ensure semantic understanding and uses immutable data layers to maintain deterministic data lineage.

Who it’s for

Robotics engineers and AI researchers who need to move from event-driven classical robotics data to the synchronous, tabular data required for Physical AI model training.

Highlights

• Zero-copy architecture: Eliminates serialization overhead for high-efficiency retrieval.
• ML module: Automates the transformation of raw, unsynchronized data into aligned, flattened formats for model training.
• Code-first approach: Provides native SDKs to query and upload data without requiring a specialized query language.
• Immutable data layers: Ensures strict data lineage and deterministic query history.