We're building data curation for Physical AI.

Every Physical AI model, from world models to humanoid robots, is only as good as the data it learns from. Labs now pretrain on millions of hours of egocentric video, then adapt to a specific robot with as little as an hour of demonstrations. Collection networks are flooding the market with footage of uneven quality.

The hard problem has moved past storing that data to knowing what's in it: what happened, what's reliable, and what will teach a model the capability you want. Today that still means paying people to watch footage, so teams check a sample and hope it represents the rest. Train on the wrong examples and the model doesn't just fail to improve. It learns the wrong lesson.

Eventual turns that footage into signals: actions over time (a failed grasp, a retry, a recovery), contact and hand–object interactions, trajectories, and whether an example is reliable enough to learn from. Every signal links back to the exact moment in the source and comes with a measure of how accurate it is. Researchers use signals to search and inspect behaviors, to select and filter training data, and as supervision targets in their own pipelines. Signals are versioned with each data release, so each dataset is better than the last.

We work with frontier Physical AI labs choosing pretraining data and with robotics teams improving their models with fleet data. It all runs on Daft, our open-source engine, which processes 2 PB of data a day at Amazon.

Our team comes from AWS, Lyft and Tesla. We powered the last generation of Physical AI in self-driving. Now we're helping the next generation decide what to train on. We've raised $30M from Felicis, CRV and Y Combinator, plus angels including the co-founders of Databricks and Perplexity.

If that's a problem you want to work on, we're hiring. If it's a problem you have, get in touch.

— Sammy, Jay, and the entire Eventual team