<!-- SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -->
<!-- SPDX-License-Identifier: Apache-2.0 -->

<a id="cts-overview"></a>

<a id="device-conformance-testing"></a>

# Device Conformance Testing

You have integrated a device that publishes hand or full-body poses into Isaac Capture. The
Conformance Test Suite (CTS) produces the evidence that the integration works: you record one
scripted take, a checker reads only that recording and returns a verdict, and you send the take
through the submission form for your pull request. This page explains the process, the two suites
and what a verdict means.

<a id="how-it-works"></a>

## How it works

**1 · Integrate**

Your device feeds Isaac Capture’s hand or body tracker, through your plugin or the
headset’s own tracking.

Add a New Device

[Add a New Device](../device/add_device.md)

**2 · Set up**

Build with Python 3.12, create the tool environments, connect the headset.

Set Up the Tools

[Set Up the Tools](setup.md)

**3 · Rehearse**

Practice the script until the performer knows every step.

[Hand](hand.md#cts-hand-rehearse) · [Full body](full_body.md#cts-full-body-rehearse)

**4 · Record**

Record one scripted take in the capture panel in your browser.

[Hand](hand.md#cts-hand-record) · [Full body](full_body.md#cts-full-body-record)

**5 · Check**

The checker reads the take and returns one verdict.

[Hand](hand.md#cts-hand-check) · [Full body](full_body.md#cts-full-body-check)

**6 · Submit**

Package the take with its report and send it through the submission form.

Submit Your Results

[Submit Your Results](submit.md)

<a id="cts-overview-suites"></a>

<a id="pick-your-suite"></a>

## Pick your suite

Run the suite for the data your device produces. A device that produces both runs both.

**Hand Preview**

- **For:** gloves, or a headset’s own optical hand tracking
- **Records:** `core.HandPoseRecord`, 26 OpenXR joints per hand, 21 of them judged
- **Script:** 11 gestures, seated at a table, about 2 minutes
- **Pacing:** automatic; a spoken cue, a count-in, a beep, then a timed hold

Hand Suite

[Hand Suite](hand.md)

**Full body**

- **For:** headset body tracking, or a mocap suit through a vendor backend
- **Records:** `core.FullBodyPoseRecord`, 24 joints
- **Script:** 10 poses and motions, standing
- **Pacing:** self-paced; you press when ready, then hold the pose or do the motion

Full-Body Suite

[Full-Body Suite](full_body.md)

<a id="what-the-checker-looks-at"></a>

## What the checker looks at

| Area                          | What it asks                                               | Examples                                                                                                                        |
|-------------------------------|------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------|
| Signal and timing             | Is the stream well formed and steady?                      | Finite values, unit quaternions, monotonic timestamps, frame rate and gaps, joint speed                                         |
| Skeleton geometry             | Is it a plausible human skeleton, in metres?               | Units, bone lengths that stay constant, human proportions, joints in their named slots<br/>and on their named sides, handedness |
| Placement and frames          | Are the joints where the body is, in the frame they claim? | Not parked at the origin, wrist not frozen, quaternion component order, positions and<br/>orientations that agree, +Y up        |
| Poses over the scripted steps | Did each step happen, in order, as labelled?               | Hands flat on a level table, pinch finger identity, arm raised overhead, squat depth,<br/>march rhythm                          |

The checker needs only the `mcap` and `flatbuffers` Python packages. It is deterministic, so
we re-run it on your take and get the same verdict. Some checks are only measured and reported;
the suite pages say which.

<a id="cts-overview-verdicts"></a>

<a id="verdicts"></a>

## Verdicts

![Verdict ladder: a device failure gives fail, else a performance failure gives retake, else
an unanswered required check or no answer at all gives insufficient_data, else pass.](../_images/verdicts.svg)

| Verdict             |   Exit status | Meaning                                                                                                          | What to do                                                                           |
|---------------------|---------------|------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------|
| `pass`              |             0 | Nothing in the recording argues against the integration.                                                         | Submit it.                                                                           |
| `fail`              |             1 | A failure blamed on the device or its plugin.                                                                    | Fix the integration and record again, or submit with an explanation.                 |
| `retake`            |             2 | The take is unusable because of how it was performed. The device is not implicated.                              | Rehearse the failed step and record again.                                           |
| `insufficient_data` |             3 | The recording cannot answer the question: too short, no channel, or a required check could<br/>not run.          | Check that the device streamed on its channel for the whole take, then record again. |
| (none)              |             4 | The check could not run: a usage error, an unknown check, or a recording or labels file<br/>that cannot be read. | Fix the command or the file named in the error, then run the check again.            |

A take checked without its labels file still gets a verdict, even `pass`, but every step check
reads `n/a`: keep the labels file beside the recording. You can submit any verdict with its
report.

### Marks in the report

- `pass`: answered, and the answer is acceptable.
- `FAIL`: answered, and the answer is not. Whom the check blames decides `fail` or `retake`.
- `meas`: measured and reported, not judged. Read the value; it is not an approval.
- `n/a`: not answered, because data was missing or a check it depends on failed. Never a pass.
- `note`: an advisory check failed. Reported; it does not move the verdict.

<a id="what-you-need"></a>

## What you need

- A Linux workstation with the repository built for Python 3.12.
- An XR headset connected through CloudXR, even when your device produces the joints.
- Your device’s plugin built and installed, if it has one.
- `aplay` (`alsa-utils`) for the spoken cues. Without it, or without speakers, the capture
  panel’s on-screen cues still guide every step.
- Room for the script (a table and chair for hand, standing space for full body), and ideally a
  second person at the screen.
- A Google account to sign in to the submission form.

Start with [Set Up the Tools](setup.md).
