Labeling
Timestamped labels on video segments.
Physical AI
Auto-label your robotics data.
Subtask labeling, detection, and 3D reconstruction.
Loved by leading
AI companies
Capabilities
Label robotics video with the primitives that matter.
Robotics video labeling using multiple models.
Compare auto-labels from different models against a shared ground truth. Line them up side by side and see how they perform.
Compare how each model performs.
The same episode, four models, one ground truth. Scored subtask by subtask, so you can pick the one that holds up on your data.
Compare models on the left. Switch camera views on the right.
0:00 / 0:53
Pick up key0:00
Alignment vs ground truth
Each model scored against human ground truth: per-frame subtask agreement, boundary overlap, sequence order, and a combined score.
| Model | Frame | IoU | Seq | Score |
|---|---|---|---|---|
| Flash | 81% | 0.74 | 0.88 | 71.2 |
| Muse | 77% | 0.69 | 0.84 | 64.8 |
| Qwen | 71% | 0.63 | 0.79 | 48.1 |
| ER-2 | 64% | 0.58 | 0.71 | 25.9 |
Interact with the objects.
Orion tracks the scene, then reconstructs each object as an interactive 3D splat you can orbit.
