lake
An open-source lakehouse for embodied-AI data. Robot fleets write immutable episode data; stateless DataFusion query nodes read exact table versions directly from object storage through Arrow Flight SQL.
The metadata layer remains a bounded catalog authority while query replicas absorb concurrent training and evaluation reads behind local caches.
Documentation
Start with the architecture, the robot-training direction, or a local
deployment. Every page is rendered from the repository's canonical
docs/ Markdown.
-
Architecture
How stateless SQL compute, bounded metadata authority, and object storage fit together.
-
Robot-training lakehouse
The product loop, data model, current capability boundary, and delivery direction.
-
Local Deploy Standards
Run the complete local stack and exercise ingest, commit, and SQL reads.