Local research tool

Train with your Skillspace footage

Check your recordings, learn a compact visual representation, and inspect its held-out results. Your files and training stay on your server.

CPU workflow · Offline by default · One training job at a time

Visual learning does not control an arm. An executable movement policy also needs aligned robot actions, a compatible robot and scene, and independent task qualification.

01Select a Skillspace

Local folder, exported ZIP, dataset path, or HTTPS link

Paths refer to files on the computer running this server. Use a local folder for footage larger than the 24 MiB upload limit; clips may be up to 128 MiB each.

Allow HTTPS intake for this source. Training still runs locally.

This server is offline. HTTPS intake is available when it starts with --allow-online.

Or upload a Skillspace ZIP. It will be saved locally and checked before training.

02Check footage and groups

Inspect usable episodes, source groups and the training/development/test split before learning. Related recordings should stay in one split. Group counts reflect declared lineage and hashes; novelty is not independently verified.

Study seed

Select your source first.

No dataset has been checked yet.

03Train locally

The first output is visual learning. The pipeline fits a small representation and tests it on held-out footage. Movement readiness is reported separately.

Check the footage before training.

No training job has started.

04Inspect and download

Held-out metrics, elapsed time and the result archive will appear after a verified run.