Two gates catch different defects
The first checks labels at full corpus scale, with no robot involved. The second checks the policy those labels produce. A defect the first misses and the second catches is the most valuable thing the loop can find.
A policy that works in the lab is a starting point. What carries it through deployment is the loop around it: evaluate, find exactly where it fails, collect the data that fixes it, retrain. grip.ai runs that loop for your robot, starting from egocentric human video.
Robot teams are not short of ideas. They are short of data that matches their robot, a way to tell whether it helps, and a cheap way to fix what breaks.
Any egocentric human video becomes training-ready trajectories for your hand: metric-scale, with contact and a per-frame feasibility score, so unusable frames are filtered before training instead of discovered after.
We define data quality as the downstream policy error it predicts, not how clean a clip looks. You get a measured number for your robot, not a sales adjective.
Failures from evaluation become targeted data requests, by verb, scene and hand. You re-collect what is broken, not everything.
Perception runs once. Each new embodiment is a cheap CPU-only retargeting pass, so a new hand is a config entry rather than a new pipeline.
Source, license class and transform chain travel with every shard, so redistributability is a queryable field.
A dataset gets you a demo. Every evaluate-and-fix iteration is what turns it into a product, and it is the part that rarely gets measured.
Human video to metric hand, camera and object geometry.
Trajectories for your robot, with feasibility scored per frame.
A vision-language-action policy on wrist motion, grasp and finger flexion.
Measured twice: are the labels right, then do they teach the policy anything?
Error by slice becomes a collection spec. Back to step 1, aimed at what failed.
The first checks labels at full corpus scale, with no robot involved. The second checks the policy those labels produce. A defect the first misses and the second catches is the most valuable thing the loop can find.
Perception is embodiment-independent and runs once. Closing the loop costs a retargeting pass and a targeted collection, not a pipeline rebuild.
We are building a neutral scorecard: for a fixed policy and embodiment, how many real robot demos reach a success threshold when pretrained on data source D.
The detail behind the outcomes above, collapsed so you can skip it.
One perception pass over human video produces an embodiment-independent artifact: hand keypoints, camera trajectory, wrist SE(3) motion and object tracks. Retargeting to a given robot is a separate, cheap pass.
Bimanual hand pose with stable per-hand identity, SLAM camera trajectory and wrist SE(3) deltas in one schema with per-frame confidence.
Scale ambiguity is resolved against the geometry of the hand itself, so actions are in meters in a world frame.
A continuous grasp signal and per-finger flexion, invariant to camera framing and hand size, and so shared between human and robot.
Human keypoints to joint trajectories for multi-finger hands, with a per-frame feasibility score alongside.
Structural validation gates a shard before it ships. Source, license class and transform chain travel with the data.
A robot is one declarative profile: backend, degrees of freedom, URDF, finger layout, joint limits. Bring yours and we add it without re-running perception.
| Profile | DoF | Role |
|---|---|---|
| leap | 16 | Primary target: inexpensive, clean URDF, widely used in academic work |
| allegro | 16 | Second dexterous hand for cross-hand generalization |
| shadow | 24 | High-DoF ceiling for the retargeting problem |
| inspire | 12 | Lower-DoF dexterous point on the curve |
| sharpa | 22 | Reference profile, retained for shard compatibility |
| gripper | 1–2 | In build — parallel-jaw floor, proving transfer across a large DoF gap |
Quality is the downstream policy error a data source's labels predict. The scorecard is built around that definition and reports curves, not single points.
Robot demos required to hit a success threshold: engine-produced vs. raw human video vs. teleop-only.
Multi-finger tasks that parallel-jaw-only pipelines structurally cannot score.
Cost and wall-clock per curated episode, published.
With and without object grounding, metric scale and feasibility filtering, so each component earns its keep in public.
We are working with a small number of teams training manipulation policies: VLA labs, dexterous-hand makers, and sim/eval vendors. Tell us what your policy gets wrong today and we will run your embodiment through the loop.
Or reach us directly at info@grip-ai.net