Introducing the Kite Platform API

Robotics teams rarely run out of ideas. They run out of infrastructure.

Every step between "we should try that" and "we know whether it worked" goes through something someone has to build and babysit: getting more data, getting GPUs, getting an evaluation you actually trust. That plumbing is where the weeks go — and it's the same plumbing at every company, rebuilt every time.

The Kite Platform exists to take that off your plate. Today we're opening the first piece of it as an API.

Starting with the data problem

Data is the bottleneck that gates all the others. A policy that only ever saw one kitchen is a policy you can't deploy, and the honest fix — collect more, in more places — costs robot time you don't have.

So the first thing on the API is augmentation. You point at a LeRobot dataset you already have, describe the variation you want in plain language, and get a new dataset back:

curl -X POST https://api.kiteml.com/v1/augmentations \
  -H "Authorization: Bearer $KITE_API_KEY" \
  -d '{
    "source": { "repo_id": "lerobot/pusht" },
    "instructions": "white marble table, warm side lighting",
    "config": { "episode_count": 10 },
    "output": { "type": "download" }
  }'

Kite regenerates the pixels, carries the original joint and action trajectory onto the new footage, and hands back a standard LeRobot dataset — either downloadable or pushed straight to your own HuggingFace account. Same robot, same motion, a world it has never seen.

If you'd rather not write HTTP, it's one line:

kite augment create --repo-id lerobot/pusht \
  -i "add a wooden floor and warmer lighting" -n 10 --wait

What that actually unlocks

The interesting part isn't that augmentation is now callable. It's what becomes possible once it is.

Data collection stops being an expedition. When augmenting means opening a browser, it's a thing you do once, before training, and then forget. When it's a call, it becomes a response to a measurement: you evaluate a policy, find the lighting condition it falls apart under, and generate data for that exact weakness — tonight, automatically, while you sleep.

Your pipeline gets a data stage. Augmentation slots in next to the steps you already run on a schedule. New demonstrations land, a job fans them out into varied scenes, training picks them up. Nobody is clicking anything.

Scale stops tracking headcount. Generating a hundred variants costs the same amount of human attention as generating one — which is none.

And because these are long, expensive jobs, the API is built for that reality rather than pretending otherwise. Retries don't double-charge you. Webhooks tell you when a run lands instead of making you poll. You can price a run before you commit to it, and cancelling refunds what was never generated.

The plan: unlock the whole loop

Augmentation is one bottleneck. There are three, and they're the same three everywhere: getting the data, getting the compute, and getting an honest verdict.

The Kite Platform is being built to remove all three, and to make them compose — because the value isn't in any one step, it's in the loop. Augment a dataset, train a policy on it, evaluate that policy across randomized simulations, see exactly where it fails, and feed that back into the next augmentation. Every team doing serious robot learning is running that loop by hand right now.

Augment is live on the API today. Training and evaluation are next. Both already work in the Kite app; what's missing is the programmatic surface, and we've built the API so they land without churn — every long-running job in Kite already reports through one shared shape, so training runs and evaluations will behave like resources you already know how to call.

The end state is the full loop over one API, with no browser anywhere in the path. We're shipping it a bottleneck at a time.

Get a key

Open the Kite app, click your name in the sidebar, and choose Developers → API keys. Pick your scopes, create the key, and copy it — it's shown once. Your first call is about ninety seconds away.

Create an API key→

Full reference, authentication, and webhook samples live at docs.kiteml.com — start with the quickstart, or go straight to the API reference.

If there's a bottleneck you want us to take on before we get to it, tell us — that's how augmentation got prioritized.