Question 1

What is Kite?

Accepted Answer

The evaluation layer for physical AI, with three services. Augment your robot datasets, Train policies on them, and Evaluate those policies across thousands of simulated episodes in a twin of your cell, with task success and a confidence interval, failures clustered by cause and by object, and every episode's video.

Question 2

What do I need to bring?

Accepted Answer

A policy, as a LeRobot checkpoint, a Hugging Face repo or an RL checkpoint, and the robot it runs on.

Question 3

Which robots and policies are supported?

Accepted Answer

Arms (SO-101, WidowX AI, YAM, OpenArm, Piper, ALOHA, UR5e), the Unitree G1, quadrupeds (Go2, Spot) and your own URDF or MJCF. The most popular policies load through LeRobot, including ACT, π0, π0.5, SmolVLA, GR00T and XVLA, alongside RL checkpoints.

Question 4

How is the twin built?

Accepted Answer

From one recorded episode. The room is rebuilt from its frames, each object is authored with working physics and checked by measurement (a cardboard box that behaves like cardboard), and the robot is placed where it stood.

Question 5

How do I get a policy to generalize to new, unseen environments?

Accepted Answer

Augment the dataset you already have: relight the episodes, move them to new rooms, add objects and distractors, with the robot's motion kept exactly. Train on the wider set, then evaluate the policy in those same conditions to see that it holds up before it meets them for real.

Question 6

Can Kite augment my dataset?

Accepted Answer

Yes. Point it at a LeRobot dataset on Hugging Face and describe the change: relight the episodes, from day to night or warm end-of-shift light, or generate new rooms, objects and distractors around them, with the robot's motion kept exactly. The augmented dataset comes back as a download or is pushed to your Hugging Face account, ready to train on.

Question 7

What does a run report?

Accepted Answer

Task success with a 95% Wilson interval, a failure histogram, success per object and per condition, the edge cases below the mean, and a video for every episode.

Question 8

Can my coding agent run evaluations?

Accepted Answer

Yes. The CLI's evaluation commands are also MCP tools on a hosted server, so Claude Code or Codex can start a run with kite_eval_run and read the report back, with nothing to install.

Question 9

Can I train on Kite too?

Accepted Answer

Yes. Start an RL run from the API on GPUs you never wire up, or fine-tune a VLA, and evaluate the checkpoint in the same place.

Question 10

How much does Kite cost?

Accepted Answer

Every account starts with 10,000 credits. Runs spend credits per episode, reserved when a run starts and settled when it ends. Enterprise: talk to a founder at raul@kiteml.com.

The evaluation layer for physical AI

Better policies, in a fraction of the time

Evaluate in your environment

Find the data gaps, close the loop

One API.
Every evaluation.

Start evaluating

Questions