Builder Guide

Mobile + XR Track — AI in your hand or headset

ExecuTorch Hackathon · October 17–18, 2026 · Founder’s Inc., San Francisco
Teams: 3–5 people · One combined competition track, with separate phone and headset build paths

Create a useful mobile or immersive experience powered by local ExecuTorch inference. Choose the path for your assigned device and connect a real input to a visible, meaningful result. You do not need to build for both devices.

Your first working demo

Mobile: camera/image input → local inference → a visible result on the phone.
XR: real image input → local inference on the headset → a visible result connected to a headset interaction. The published XR starter uses DeepLab V3 image segmentation.

Suggested directions: a private visual assistant, an image-understanding interaction, or an immersive experience that responds to model output. Scope camera access and interaction APIs early; a command-line inference demo alone does not provide the finished application.

Get started

1. Prepare your device and host. Complete the loaner agreement, enable debugging, and authorize Android Debug Bridge (adb). The XR walkthrough uses Linux/WSL. Use Android development tools when building your app.
2. Run the matching starter. For Mobile, obtain the S25 model/runtime bundle and launch instructions from the track team; these are not in the repository. The XR ZIP contains the QNN model, runner, and libraries, but is missing its advertised launch script and checksum file. Ask for the corrected launcher, or adapt the guide’s manual QNN staging commands to the ZIP’s flat file layout. Source builds are optional.
3. Check XR firmware first. The guide reports NPU validation on I610UEU1AYIA; I610UEU2AZF3 blocks NPU access. For blocked units, obtain the separate XNNPACK CPU model and runner; they are absent from the XR ZIP. Other firmware needs device verification. Phone and headset artifacts are not interchangeable.
4. Build and verify the experience. Match preprocessing to the model, display its output, and test multiple inputs plus an app restart. Record end-to-end latency, memory, and sustained battery/thermal behavior on the actual device.

App integration dependency: the guide’s Android Archive (AAR) / Module integration has not been validated on Galaxy XR. Obtain a backend-compatible app runtime and prove integration early. Keep each model with its matching QAIRT libraries; the documented XR bundle uses ExecuTorch v1.5.1 and QAIRT 2.51.0.260929. A successful default-input run does not establish real-image accuracy.

Submission and judging

Submit by Sunday, October 18, 1:00 PM Pacific through the sponsor-announced submission form. Provide a public GitHub repository and demo video, an open-source license, and a README covering the application, team names/emails, dependencies, setup, and usage. The app must install and run on its intended device, use ExecuTorch, and keep most components local; hybrid cloud use is allowed. Volunteer live demos have a five-minute limit.

Scoring / 100: use case and impact 40 · technical implementation 25 · on-device AI and resource efficiency 15 · deployment, documentation, and accessibility 10 · presentation 10. Mobile and XR share one track winner.

Help: Mobile + XR contact listed in the event guide: Emre. Participant Discord.

Sources and next steps: Mobile + XR guide · Event guide. Reviewed October 16, 2026; firmware guidance follows the track guide, which conflicts with the repository overview.