Every strong policy starts with great data

We spent the last few years turning a large part of the internet into some of the most widely used open LLM pre-training datasets. That work gave us a front-row view of what happens when a field learns to collect, evaluate, and process data at scale: progress gets very fast.

Robotics is reaching a similar point. More people are moving from pure software into the physical world, and the teams that learn fastest from their robot data will have a real advantage. Robot fleets and wearable cameras can already produce more recordings than teams can feasibly review. The temporal structure, physical motion, outcomes, and quality signals needed for training are still buried inside raw episodes.

Macrodata Labs exists to help robotics teams turn those raw episodes into progress. Demonstrations, frames, trajectories, annotations, and sensor streams should be easier to inspect, enrich, and turn into better policies. A dataset becomes useful for training when the signals inside it are extracted consistently.

Our pipelines do that extraction at scale. They segment long robot and egocentric recordings into timestamped subtasks, recover metric 3D hand motion from moving cameras, assess whether recordings contain the signal a team needs, and estimate task progress and reward signals. Each pipeline is adapted and evaluated on representative data before being run across the wider collection.