You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

ARE Agent Studio — Signal Control Room

ARE Agent Studio VLA Demonstrations

SIGNAL CONTROL ROOM · Publication-approved human demonstrations with provenance, not synthetic showcase data

This Dataset Card describes the private provenance repository for a future ARE Agent Studio release. It does not assert that any raw samples have been publicly released or that an APK is available.

When a dataset release is actually published, the release record must name the exact Hub repository and revision and bind them to the dataset_manifest.json SHA-256. Do not replace this statement with a guessed or placeholder Hub URL.

Publication boundary

The public export pipeline includes only rows marked publication.allowed=true with publication.basis=user_confirmed at capture time. Recording a sample locally does not imply permission to publish it. Operators remain responsible for having appropriate rights to distribute captured game imagery and other content.

Signal lane Required evidence before a Hub release
Capture A complete visible frame and an observed human action are paired locally.
Consent The individual row carries explicit publication.allowed=true and publication.basis=user_confirmed.
Snapshot The builder creates JSONL, episodes, decoded frames, and a manifest with SHA-256 values.
Publish The publisher re-hashes the snapshot immediately before upload and refuses an empty or manipulated export.

The artwork above is a project visual identity. It is not a captured frame, a device readback, a benchmark visualisation, or a sample contained in a dataset snapshot.

Schema

The current schema is are-agent-vla.v1. The snapshot also contains data/episodes.jsonl, which groups approved samples by capture session without inventing missing steps. Important fields include image, source (human_demo or dagger_correction), normalized target_action_chunk, a 16-dimensional observed visual feature vector when available, state values plus state_mask, action metadata, client metadata, and publication metadata.

Unknown state values are represented by a neutral numeric placeholder only together with state_mask=0; they must not be interpreted as observed zeros.

Provenance

The export contains dataset_manifest.json, including the SHA-256 of the source append-only ledger, the exported JSONL, every decoded frame, and the manifest itself. The local dataset daemon independently content-addresses samples and returns a receipt when rows are accepted.

Intended use

Research and prototyping around behavioral cloning, DAgger-style human correction, visual control policies, dataset tooling, and reproducible agent evaluation. The dataset does not claim that an action was executed on Android merely because it appears as a policy prediction; device execution is a separate ADB-acknowledged path in ARE Agent Studio.

Access and non-claims

This source card supplies no raw-dataset download URL, APK download, automatic-action success count, or performance claim. The private provenance repository is not linked as a public dataset release; direct dataset retrieval is never served by the ARE public metrics endpoint. Access settings, license selection, and release visibility must be decided by the repository owner for the published snapshot, not inferred from this card.

License note

No dataset license is asserted by the automation. The repository owner must select a license compatible with the rights in the captured imagery before the first public release.

Downloads last month
18