# Train your first policy

A policy is the model that controls the robot. You train it in simulation, publish it to
your Hugging Face account and enter it here, all from a workshop on your machine. You do not
need a physical robot.

## The workshop

Open a challenge and copy the line under **Train my own robot** into a terminal. It clones
[microduck-challenges](https://github.com/pollen-robotics/microduck-challenges) and opens the
workshop in your browser, on that challenge:

1. **Ready to train?**, on the workshop's home, says what is set up: Hugging Face Jobs billing,
   your fork of the challenges, an NVIDIA GPU. Log in to Hugging Face with the button at the top
   right: it opens huggingface.co, where you approve.
2. **New training**, on the challenge's page: change the challenge's files if you like, commit them
   (the workshop asks), choose Hugging Face Jobs or this machine, and train. Each training gets a
   ticket; several can run at once.
3. **Every checkpoint is evaluated** with this Arena's rules, on your machine, as it arrives. The
   time is indicative: only the Arena times an entry.
4. Open a ticket, pick a checkpoint, and **Publish and compete**: the policy goes to your Hugging
   Face account and is entered here, with a word for the commentators, and its result opens.

## From a shell

Everything the workshop does is a command you can run yourself in a clone of
microduck-challenges (the challenge's `task` is in its `challenge.toml`):

```bash
uv run train Mjlab-Sprint2m-MicroDuck --env.scene.num-envs 4096 --agent.max-iterations 3000 \
    --agent.seed 1 --agent.logger tensorboard --agent.run-name first
# no NVIDIA GPU: add --hf-jobs --no-wandb, then bring the run back with
# uv run hf download <the checkpoint repo it printed> --local-dir logs/rsl_rl
uv run publish --run logs/rsl_rl/sprint_2m/<the run> --repo <you>/microduck-first --no-private --enter
```

`--enter` races it here and prints its time and its page. Add `--livery purple` for a
colour, and `--speed 0.6` to be timed at one speed instead of the sweep. The Arena expects 61
observations, 14 actions and a file no larger than 64 MiB; `publish` checks all three.

## Already have an `.onnx`

Race it on your machine with this Arena's rules, in the same clone:

```bash
uv run --group workshop arena score policy.onnx --event sprint-2m
```

To enter it, publish it to your Hugging Face account with the event it is for:

```bash
uv run publish --onnx policy.onnx --repo <you>/microduck-<name> --kind perpetual --no-private --enter --event sprint-2m
```

For a roller challenge, add `--accessories rollers`.