Health checks for robot learning datasets
Kinesift finds the problems hiding in teleoperation and demonstration data — out-of-sync streams, dropped frames, frozen states, duplicates, broken metadata — before they cost you training runs.
Request a health checkWhat we check
Synchronization
Image and action streams out of step; cameras out of step with each other.
Dropped frames
Gaps in timestamps and in the video itself.
State and action sanity
Values outside joint limits, implausible values, sudden jumps, NaN or infinite values.
Frozen or static episodes
Episodes where nothing moves, or where the recorded state does not follow the commanded action.
Gripper signal
Gripper values that barely change across an episode.
Length and content
Abnormally short or long episodes, empty episodes, all-black video.
Duplicates
Exact and near-duplicate episodes.
Metadata
Episode counts, indices, task tables, fields, fps and missing files that disagree with the data.
Supports LeRobot dataset formats v2.0, v2.1 and v3.0. Every finding in a report comes with the exact command to reproduce it.
How it works
- Public dataset: send us the Hugging Face dataset ID. We run the check and send you a report.
- Private dataset: you run our checker on your own machine and share only the report. Your data never leaves your infrastructure.
- You get a report listing each issue, the affected episodes, how to reproduce it, and a suggested fix.
Founding-customer pricing
One-time health check
Per dataset, depending on size. Delivered within 5 business days. You pay after delivery — if we find nothing you can fix, there is no charge.
Ongoing checks
Every time your dataset is updated, it is checked again automatically.
Does filtering data actually help?
We are running a pre-registered study: does removing low-quality episodes improve policy success more than removing the same number of episodes at random? We will publish the results either way.