Datoric

Teams building physical AI and robotics models need large volumes of real world data covering video, motion, and tactile signals, but sourcing this data in a licensed and ethically collected way is difficult. When a model fails or falls short of a capability target, teams often lack a systematic way to trace that failure back to a specific data gap. Without controlled experiments across modalities, it is hard to know which kind of data collection would actually fix the problem before committing to a full production dataset.

Datoric offers existing datasets across modalities including egocentric video of household and workplace tasks with task, action, and object labels, UMI bimanual demonstrations with gripper views, end-effector poses, and gripper width, and teleoperation data with synchronized video, states, actions, and robot and sensor configuration. Beyond off-the-shelf datasets, Datoric runs a research process that starts from a model failure or capability target, formulates a research question, runs controlled collection experiments across video, motion, and tactile modalities, and evaluates results with internal systems that estimate treatment effects. Only after a method is shown to transfer in downstream evaluation does Datoric specify it for production-scale data collection.