# What MouseMouse is

> The four parts, Train, Inference, Datasets and Policies, and how they close one loop from data to robot and back.

Prompt for your AI: Read https://mousemouse.ai/docs/what-mousemouse-is.md first. Explain what MouseMouse is and how its four parts fit together, then tell me where I should start. Ask me before anything that costs money or moves a robot.

MouseMouse is the open infrastructure for Physical AI. It is a hub where
robots, datasets and policies are published and reused. It is also the compute
to train and run them. This page explains its four parts and how they form one
loop.

## The four parts

**Train.** A training run takes a robot and a template, and produces a policy.
Reinforcement learning trains a walking or balance policy against the robot's
own simulation model. A vision-language-action policy is fine-tuned on a
dataset of recordings. A run trains on cloud GPUs or on your own machine, and
tests every checkpoint in simulation as it goes. [Train](/train) lists the
templates, where they run and what they cost.

**Inference.** A trained policy reaches a robot in one of two ways. A small,
fast policy is deployed to the robot and runs on its own computer. A large
policy is served from a GPU, and the robot asks it for its next actions. In
both cases a person approves the run, and a person can stop it.
[Inference](/inference) compares the two paths.

**Datasets.** Recordings in the LeRobot format. Frame by frame, they hold what
the robot saw, what it measured and what it was told to do. Upload your own, import
one from Hugging Face, or record one by running a policy on a real robot.
Every dataset has an episode viewer. [Datasets](/datasets) lists them, and
[Dataset format](/docs/dataset-format) says what the hub reads.

**Policies.** Every policy on the hub names the robot it was trained for and
its inputs and outputs. It also carries where it came from and its sim test
results. Policies
from Hugging Face are listed beside the hub's own. Anyone can download a public
policy and test it in simulation. [Policies](/policies) lists them.

## How they close the loop

The four parts are stages of one loop, and each stage feeds the next.

1. A dataset, or a simulation, goes into a training run.
2. The run produces a policy, with a sim test of each checkpoint.
3. A person approves the policy onto a robot, and the robot runs it.
4. The robot records what happened: a signed trail of the run, and the logs it
   collected.
5. Those recordings become a dataset, and the next run starts from them.

The Training Coach reads a finished run and proposes the next experiment, with
every claim cited. The [Coach](/coach) page shows an example.

## What stays open

A robot is an MJCF or a URDF file with its meshes, checked by the hub before it
is published. A dataset is a LeRobot dataset. A policy is an ONNX file or a
LeRobot checkpoint, with the robot it fits written beside it. Public robots,
datasets and policies can be read, downloaded and tested by anyone.

One API serves the web app, the `lucen` command line and the Python package.
The MCP server for your own AI uses the same API. The [API reference](/docs/api) lists every
operation. You can also [run your own hub](/docs/self-host).

## Where to go next

- [Train and sim-test your first policy on Laika](/docs/first-policy), the
  tutorial.
- [How approvals keep your robot safe](/docs/approvals), the explanation of
  what happens before a policy runs on a robot.
- [Pricing](/pricing), for what runs free and what the cloud costs.