> ## Documentation Index
> Fetch the complete documentation index at: https://docs.dreamscalelabs.com/llms.txt
> Use this file to discover all available pages before exploring further.

> ## Agent Instructions
> Treat the Dreamscale CLI and Python SDK as the supported public integration surfaces.
> Prefer `dreamscale run <robot>` for supported robots and `dreamscale.connect()` with a context manager in Python.
> Do not infer or recommend internal control-plane APIs.

# HUD

> Record and grade real-robot runs on HUD, with camera video, joints, and actions for every task.

[HUD](https://hud.ai) records each robot task as a trace you can replay and
grade. Add `--hud` to any `dreamscale run` command.

## 1. Connect HUD

```bash theme={null}
dreamscale integrations hud setup
```

This installs HUD support, opens HUD's API key page, and saves the key you
paste. You only do this once per computer.

## 2. Record a run

```bash theme={null}
dreamscale run yam --model molmoact2-bimanual-yam --hud --hud-job "Towel folding"
```

`--hud` works the same with `dreamscale run so101`. The run itself doesn't
change.

* Each `dreamscale run` session is one HUD **job**. `--hud-job` names it.
* Each task is one **trace**, with the instruction, one video per camera,
  measured and commanded joints, and the model's action chunks.
* Tasks with the same instruction are grouped, so the job shows a success rate
  per instruction.

## 3. Grade each task

When a task stops, the robot holds its final pose and asks for a grade:

| Answer | Grade |
| - | - |
| `y` | Success |
| `n` | Failure |
| A number from 0 to 1 | Partial score |
| `n dropped the towel` | Grade plus a note |
| Enter | Leave ungraded |

The grade becomes the trace's reward on HUD.

<Note>
  Recording never slows the robot. If it falls behind, it drops recorded frames
  and the trace reports how many. `--hud` can't be combined with
  `--inspect-robots`.
</Note>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.