Actions and segment attributes, placed in the context of each robot episode. Turn the recording into a structured sequence of events.
The Data infrastructure for Physical AI, Dexterous Manipulation, General Purpose Robotics, World Action Models, and VLAs.
Data for Physical AI
Ground Truth Annotations for Robotics Data
Hyphenbox turns raw episodes into densely annotated discrete subtasks with failure spans detected in each step for evaluating robot policies faster
Dense subtask annotations and failure spans for your robot episodes.
Move over the arm to explore its motion. With keyboard focus, use left and right arrows to turn, up and down to move the gripper, and Escape to return to rest. Motion follows your system preference.
Backed by roboticists from
The world is rich in motion.
Make it rich in meaning.
Turn physical interactions into structured data for robot learning. Explore the outputs for each type of recording.
TeleoperationYour recordings.
Robot or egocentric video.
Hyphenbox
Read productive steps, waiting, and recoveries alongside the source footage, with reward signals that describe task progress.
Failures marked as intervals, with a clear start and end. Inspect failed grasps, dropped objects, and the surrounding action.
The work, up closeReal tasks.
Real tasks.
Real-world complexity.
From a circuit board to a warehouse floor.
Explore physical intelligence in the making.
01 / 05
Small components. Intricate gestures. Every interaction in view.
Reach, lift, place. Follow the motion behind everyday work.
Dexterous manipulation in an unstructured environment.
Fine hand motion across a careful, repeatable task.
Hands, tools, and trajectories through skilled physical work.
Each sample includes source video and reconstructed hand motion.Explore all recordings
Built around your dataYour task.
Your task.
Our starting
point.
Every dataset has its own context.
We start with yours.
An assembly study
