Figure Index collects human-recorded video for robot learning. A clip can show how a task unfolds without supplying the joint commands, forces, or synchronized robot state needed to reproduce it. For a robotics team, the practical question is what supervision remains after a clip passes through the collection and annotation pipeline.
On August 25, 2026, Figure described Index as a Figure-exclusive pipeline feeding its Helix program. The company says uploads pass through filtering, fraud review, deduplication, rebalancing, and annotation. Those stages make Index more than a raw upload bucket, but the official pages checked on September 20 do not publish acceptance rates, the final training-set size, or an independent evaluation that isolates the lift from Index data. This is a review of those disclosures; I did not use the app or inspect its private dataset.
Human video and robot demonstrations teach different signals
A human video can show objects, contact sequences, room layouts, tool use, and many ways to complete a task. That coverage can support representation learning, task understanding, captioning, or planning. A supervised robot demonstration adds a different layer: the observations are synchronized with actions executed by a known embodiment.
| Question | Acquired human video | Supervised robot demonstration |
|---|---|---|
| What is directly observed? | Human motion, objects, scenes, contacts, and outcomes visible to the camera | Robot observations plus the actions or commands issued during the episode |
| What may be missing? | Joint state, gripper state, force, calibration, action timing, and robot failure labels | Human task diversity beyond the robot setup and collection protocol |
| What can it teach most directly? | Visual concepts, task order, object interaction patterns, and environmental variation | A policy how a specified robot maps observations to executable actions |
| What must be checked? | Capture quality, permission, task meaning, duplicates, coverage, and annotation | Synchronization, action semantics, embodiment match, success criteria, and split integrity |
The boundary is not absolute. A pipeline can infer pose, add captions, or align video with other signals. The point is to name those transformations. Video volume alone cannot establish that the resulting examples contain the control targets a robot policy needs.
Index filters the acquisition stream before annotation
Figure’s Index announcement gives a concrete processing sequence. Automated checks screen technical, visual, and semantic quality. Human analysts audit samples for attempts to evade the filters. Video segments above a similarity threshold to accepted data are discarded. The remainder is rebalanced with task quotas and embedding-based clusters, then assigned hierarchical text captions.
Consider an illustrative pair of submissions showing the same countertop, object, camera angle, and near-identical action. Two uploads increase the acquisition counter, but the second may add little new supervision. A different grasp, failure, viewpoint, or environment may be more valuable even if its clip is less polished. Figure discloses the logic of deduplication and balancing, not the undisclosed threshold values or rejection rates, so this example explains the mechanism rather than estimating its yield.
Five questions define whether a clip helps a robot
- What is the learning target? A visual encoder, task planner, reward model, and low-level controller need different labels.
- Does the viewpoint expose the decisive event? Occluded contact or an off-camera outcome can make a clean clip ambiguous.
- Is the example distinct? Count variation in tasks, objects, environments, strategies, and failures after deduplication.
- Which embodiment signals are absent? Identify the action, timing, proprioception, and force information that must come from robot data or another labeling step.
- What evaluation can attribute value? Hold the model, robot tasks, and test split constant while changing the added data source.
This sequence is specific to interpreting Index as a source for robot learning. A broader operational checklist appears in the robot learning data-quality audit. For a contrasting example with synchronized robot demonstrations and separately collected corrections, see the XR-2 data review.
Exclusive access and model evidence remain separate questions
The public Index page offers access to the contributor app. It does not expose a dataset download, API, catalog, research license, or reproducible training package. Figure’s announcement calls the pipeline exclusive, so readers should not treat the video collection as an open dataset.
Governance also belongs inside the data-quality boundary. Figure’s current privacy policy classifies photos, videos, and recordings of people or their environments as sensory data. A dataset assessment should therefore ask how capture permission and downstream use are handled. The policy alone does not establish permission for every clip or every intended training use.
Index demonstrates a high-throughput acquisition and curation system described by its operator. The inspected pages do not provide an independent audit or controlled result that isolates how much Index improves Helix. The sound conclusion is therefore conditional: crowdsourced human video can expand task and environment coverage after filtering and labeling, while robot-action supervision, access rights, and measured model lift still require separate evidence.