Unstructured task sequences
Continuous demonstrations lack tasks, subtasks, actions, and atomic actions.
UMI Data
Turn UMI demonstrations into stable, reusable training data
UMI captures rich human demonstrations in real environments. Data still needs governance, annotation, and quality validation across operators, tasks, and scenes before it can enter policy-learning workflows reliably.
01
Data we handle
- UMI device-camera video
- Single-hand and bimanual demonstrations
- Gripper pose and trajectories
- Gripper open/close state
- Tactile information when available
- Task labels, objects, and environment metadata
02
Common UMI data problems
Inconsistent action boundaries
Operators interpret action start and end differently.
Vision and operation semantics diverge
Video lacks structured objects, actions, state changes, and goals.
Failures and retries are unlabeled
Stops, retries, recovery, and ineffective actions affect sample quality.
03
Annotation and governance
- Hierarchical task decomposition
- Action semantics and object relationships
- Gripper state, trajectory, and outcome annotation
- Success, failure, interruption, retry, and recovery labels
- Task completeness, trajectory anomaly, and batch consistency QA
04
Suitable training tasks
- Robot manipulation policy learning
- VLA model training
- Long-horizon task understanding
- Human-demonstration transfer
- Bimanual coordination
- Success and failure modeling
FAQ
Frequently asked questions
How is UMI data different from ordinary egocentric video?
UMI data typically adds device-camera video, gripper pose, trajectories, and open/close state. Depending on the device, it may include tactile signals, so device view, motion, gripper state, touch, and the full task must be interpreted together.
Can KeenTruth use an existing task vocabulary?
Yes. We validate vocabulary coverage, hierarchy, and boundary rules through a pilot.
Are failed and retried demonstrations removed?
Not by default. We label, classify, or separate them based on the training objective.
Do you support bimanual tasks?
Yes. We annotate action phases, object relationships, and coordination in bimanual tasks.
Contact
Move raw embodied data into training faster.
Tell us about your data, annotation requirements, and quality expectations.