DatoricUMI Dual-Camera Annotated Manipulation Dataset
500 hours · bimanual demonstrations
Train robot manipulation policies from bimanual UMI demonstrations that pair left and right gripper views with pose, gripper width, task structure, and outcomes.
Dataset specification
- Dual wrist-mounted GoPro video with embedded accelerometer and gyroscope
- RGB observations, 6-DoF end-effector pose, and gripper width
- Aligned observations and telemetry at 10–20 Hz
- Bimanual clock and frame matching within 1/60 second
- Zarr episodes with task, object, success, and subtask labels
Intended uses
- Robot manipulation learning from gripper demonstrations
- Bimanual coordination and action prediction
- Task-conditioned policy training
Scope notes
- The selected tasks, demonstration subset, and delivery scope are confirmed during review.
- Target robot action conventions and training requirements are agreed for the intended use.
Collection scope
Demonstrations
- Bimanual tasks performed by consented operators using instrumented grippers
- Left-hand, right-hand, and coordinated bimanual execution
Capture
- Dual wrist-mounted GoPro video
- Embedded accelerometer and gyroscope telemetry
Catalog volume
- 500 hours
- Task mix and the selected delivery subset confirmed during review
Delivery
- Zarr episodes with RGB, 6-DoF end-effector pose, and gripper width aligned at 10–20 Hz
- Episode boundaries, subtask actions, and success labels
Annotation & metadata fields
- Task and object labels
- Hand and manipulation mode
- Left-hand, right-hand, and bimanual execution
- Success and episode boundaries
- Subtask action labels
Capture methodology
- Consented operators perform manipulation tasks with instrumented grippers
- Left and right camera views are paired with gripper telemetry
- Clock and frame matching within 1/60 second
- RGB observations, end-effector pose, and gripper width aligned at 10–20 Hz
Provenance & rights chain
- Recorded by consented participants or operators
- Licensing and the selected delivery scope are reviewed with Datoric
Quality, duplicates & PII
How submissions are reviewed and cleaned before they are accepted into the dataset.
- SLAM and calibration review
- Dropped-episode checks
- Trajectory integrity and bimanual alignment checks
Formats & delivery
- Dual GoPro MP4 video with embedded inertial telemetry
- Zarr episodes with aligned observations, poses, and gripper width
- Episode and subtask annotations
Rights & license scope
Licensed by Datoric for robotics training, with final scope governed by the agreement for the selected delivery.
Frequently asked
- What is included in a UMI episode?
- Left and right gripper views are paired with RGB observations, 6-DoF end-effector pose, gripper width, task and object labels, success, and episode boundaries.
- How are the two hands aligned?
- The catalog specifies bimanual clock and frame matching within 1/60 second, with observations and telemetry aligned at 10–20 Hz in Zarr.
- Can we review a task-specific subset?
- Request a review package and share your task mix, target embodiment, and training schema so the proposed subset can be scoped for your pipeline.