Data
Data collected for the conditions your system will meet.
Capture, annotation, evaluation and delivery, run as one process with one record of how every frame was made.
01 Capture
Multi-sensor, synchronized, calibrated.
We record on your hardware or ours, and log everything needed to trust the timestamps.
| Modality | What we capture | Typically used for |
|---|---|---|
| RGB cameras | Fixed and wrist-mounted, with calibrated intrinsics and extrinsics. | Detection, segmentation, policy learning |
| Depth | Stereo and time-of-flight depth, aligned to RGB. | Grasping, navigation, reconstruction |
| LiDAR | Spinning and solid-state point clouds with motion compensation. | 3D detection, tracking, mapping |
| Thermal and IR | Long-wave infrared, radiometric where the sensor allows. | Low-light and obscured perception |
| Radar | Automotive-class and short-range radar where the platform supports it. | Perception through dust, smoke and rain |
| Inertial and state | IMU, joint state, odometry and GNSS where available. | Sensor fusion, control, policy learning |
02 Annotation
Labels with a schema, a guide and an owner.
- 01
Boxes and masks
2D boxes, polygons and instance masks, with visibility and occlusion attributes.
- 02
3D cuboids and tracks
Consistent object identities across frames, sensors and sessions.
- 03
Semantic segmentation
Dense labels for scenes where a single class per pixel is the right answer.
- 04
Keypoints and pose
Articulated bodies, tools and parts, with confidence flags where views are limited.
- 05
Grasps and affordances
Contact points and approach directions for manipulation datasets.
- 06
Language and task labels
Instructions, sub-task boundaries and success or failure outcomes.
03 Evaluation
Benchmarks you can trust to stay put.
Held-out sets are built to your deployment conditions and versioned, so a score today means the same thing next quarter.
Training data tells a model what to do. An evaluation set tells you whether it works. We build them separately, with different collection sessions and different annotators, so nothing leaks between the two.
- Condition coverageSliced by lighting, weather, clutter and range, so you see where it fails.
- Frozen splitsDefinitions are versioned and never edited in place.
- Error reviewDisputed labels are adjudicated and documented before release.
04 Delivery
Formats your pipeline already reads.
| Layer | Formats |
|---|---|
| Sensor logs | MCAP, ROS bags, or raw sensor files with sync metadata. |
| Episodes | Parquet, HDF5 or RLDS-style records with observations, actions and outcomes. |
| Annotations | COCO-, KITTI- or nuScenes-style JSON. Custom schemas on request. |
| Documentation | Label guide, calibration files, split definitions, QA report and changelog. |
Example
What a specification looks like.
A scoping document is the first deliverable. This is the shape of one.
- Task
- Mobile manipulation in cluttered indoor scenes
- Sensors
- 2x RGB (head, wrist), stereo depth, 32-beam LiDAR, IMU, joint state
- Capture
- Hardware-triggered, synchronized, calibration logged per session
- Labels
- Instance masks, 3D object poses, grasp points, sub-task boundaries, outcome
- Conditions
- Lighting variation, partial occlusion, distractor objects, two floor surfaces
- Splits
- Train, validation, held-out evaluation from separate sessions
- Delivery
- MCAP and Parquet, COCO-style annotations, calibration files, QA report