Fleet data you can't search is data you can't train on.Search your whole fleet.
Every episode your robots record, indexed by what actually happened, and one question away from a training set.
The multimodal data bottleneck
Logs aren't a dataset
Recorded, then lost
Fleet logs pile up in buckets as bags and videos named by date and robot. Finding the twenty episodes you need means scrubbing hours of footage.
Unusable without context
An episode with no task, no step boundaries, no outcome and no sensor health can't be filtered, so it never makes it into a training set.
A moat that leaks
Real-world data is the advantage competitors can't buy, but only if you can find it and use it again. Unindexed, it's storage cost.
Not a mock-up. 2,131 real episodes, 8 public datasets.
Watch a fleet get indexed
Every tile is one episode from a public robot dataset on Hugging Face, run through Annoroid's automated checks. Watch it go from raw logs to a searchable index, then ask your own questions and export the answer.
Ingesting episodes
Ask your fleet…
- Episodes ingested
- 0
- Checked
- 0
- Training-ready
- 0
- With a problem
- 0
- Timing gaps0
- Frozen sensor0
- State spikes0
- Missing values0
- Idle over 1 s0
- Camera out of sync0
- Joint drifts from command0
- Unusually short0
Now run it on yours.
Free fleet data health report
Point us at a sample of your fleet logs. In 48 hours you see what the demo shows, on your own robots: what's training-ready, what isn't and why, and the answers to the questions you care about.
- Every episode in the sample indexed and checked, like the demo above
- Training-ready share by robot, site and task, with the problems that block the rest
- A training set built from a question you choose, exported in your format
- Yours to keep, whether or not we work together
From report to contract
- 1Agree what good looks likeBefore we start, we write down the questions and acceptance criteria together, so the result is judged on your terms.
- 2Report in 48 hoursYour sample comes back indexed, checked and verified, with the training set you asked for and the evidence behind every flag.
- 3Decide on your own numbersA review call on your data, not a sales deck. If it doesn't meet the criteria, we redo it or you walk away with the output.
- 4Production, in writingPer-episode pricing, turnaround and acceptance criteria in the contract, with free re-work when we miss them.
One observation pipeline
for every robot, every site, every run.
- 01
Connect the fleet
Read-only access to your bucket: AWS S3, Google Cloud Storage, Cloudflare R2 or any S3-compatible storage. New logs are picked up as they land.
- 02
Every episode indexed
The Annoroid pipeline runs on each one: quality checks, sensor health, step segments, task language, outcome and failure type, verified by people.
- 03
Ask, then curate
Search by what happened, not by file name. Save any question as a training set that updates itself when new matching episodes arrive.
- 04
Export and retrain
Pull a set as LeRobot, HDF5 or a Parquet manifest. Failures found in production feed the next round, and the loop closes.
What the fleet gets wrong in the field becomes the next training set. The loop closes itself.
Every episode carries
The index
The same outputs as the Annoroid pipeline, kept for every episode, joined into one table you can query.
| Episode | Robot, site, date, duration, source format and file location |
|---|---|
| Quality | Overall score, and every check with the time ranges it points at |
| Sensor health | Dropped frames, frozen streams, camera and joint-state sync drift, exposure and blur |
| Task | The episode's instruction, step boundaries and a short instruction for each step |
| Outcome | Success or failure, the frame where it failed, and the failure type |
| Contents | Labeled objects and hand-object contacts, where your project includes them |
| Similarity | An embedding per episode and per step, for “find more like this” |
Centralise observation.
Compound the moat.
One pipeline, every robot
The same schema across robots, sites and data types, so a question asked once works fleet-wide.
It compounds
Every new episode is indexed on arrival. The more the fleet runs, the more each question can answer.
You own it
The index lives next to your data, as Parquet in your storage. No lock-in: take it anywhere.
Verified, not guessed
Segments, outcomes and failure types are checked by trained reviewers, so a filter means what it says.
Early access: Annoroid Index runs on the same pipeline as every Annoroid job, with storage connected during onboarding. Search and saved sets arrive in the dashboard and the API as the early-access programme grows.
Free pilot · 48 hours
Index a pilot batch.
Send 100 episodes from your fleet. You get them back indexed, plus a training set built from a question you choose. Free.