Eventual

Data Curation for Physical AI

Eventual builds data curation for Physical AI. We turn millions of hours of video and sensor data into the signals that decide what the next generation of models trains on.

We process footage at petabyte scale and turn it into signals. Every signal links back to the exact moment in the source and comes with a measure of how accurate it is. There are four families:

• Actions - What happened and in what order: a failed grasp, a retry, a recovery.
• Contact and hand–object interactions - When a hand touches an object, how long it holds it, and which hand supports while the other works.
• Trajectories - How hands and objects move through the scene, frame by frame.
• Quality - Whether an example is reliable enough to learn from, or will teach the model the wrong lesson.

Labs use these signals to search, select and filter the data that goes into each training run, and as supervision targets. Signals are versioned with each data release, so each dataset is better than the last.

We're also the company behind Daft, the distributed multimodal data engine that makes it easy to run models on your Images, Audio, Video, and Robot Logs.

eventual.ai · blog · github · careers · get in touch