Platform · Data

Every run is a dataset

Not a clip and a checkmark. The exact policy inputs, every command and its receipt, timing, the trajectory and an independent grade, in one archive you can replay, hand to someone else or train against.

Record a run Downloads in the docs

If it happened in the run, it is in the archive. If it is not in the archive, we do not claim it.

evaluation.zipeverything the run produced

evaluation.zip
├─ README.md          what is inside and how to replay it
├─ manifest.json      model, sensors, controller, scene, budgets, source hashes
├─ metrics.json       success, termination, steps, seconds, returned models
├─ observations/      the exact images and state the model received
├─ wire/              provider requests and responses
├─ commands.jsonl     every command, its receipt and revision
├─ model.mjb          the compiled simulator model
├─ trajectory.npz     recorded state at 20 Hz
└─ replay.py          recorded-state replay helper

What is in the archive

A run finishes by writing one archive. It holds what the model saw, what it sent, what came back, what the arm did and what the simulator concluded. The download is not made available until the archive has closed, so a partial file never looks like a result.

Spectator view and the model’s own camera frames, stored separately and labelled.
Command receipts with identifiers and revisions. A replay can show it ran the same actions.
Provider requests and responses with the model IDs the provider returned. Timing and token usage per request.

commands.jsonlone line · shape from the current release

{
  "command_id": "…",
  "name": "move_to",
  "arguments": { "x": 2.665, "y": 0.100, "z": 0.945, "open": 1, "steps": 28 },
  "expected_revision": 254,
  "before_revision": 254,
  "after_revision": 282,
  "note": "Inspect the achieved descent and whether the
           fingertip pads straddle the mug body."
}

Values from program 9 of the kitchen run.

receipts

Every command, receipted

Each primitive call is written to the command log with the revision the model expected the world to be at, the revision it found and the revision it left behind. The note the model attached travels with it. Replays check the same receipts, which is how a replay can prove it ran the same actions rather than merely looking similar.

Command identifiers, expected and achieved revisions, arguments and notes.
Replays are compared receipt by receipt, not eyeballed.
Each command carries a digest of its name, arguments and expected revision.

provenance.jsoneveryday kitchen · public file

{
  "id": "kitchen",
  "author": "Gnangarra",
  "license": "CC BY 2.5 AU",
  "sha256": "f71a409d…638cac",
  "changes": "Reference resized for display; 3D reconstruction
              is a separate generated approximation.",
  "version": "0a7f291999595c9db1e710adf19eb136",
  "license_applies_to": "reference and distributed reconstruction",
  "generator": "Atom using the deployed RobotGym workflow"
}

provenance

Frozen before it starts

The manifest freezes the model, sensors, controller, scene version, budgets and source hashes before the first control step. If anything changed between two runs, the manifests say what. Source photographs carry their author, license and SHA-256 in the provenance record, and the harness commit is pinned in the evidence.

Scene versions are content-addressed. The kitchen run records scene version 0a7f2919 and scene hash 1d824047.
Returned model IDs are written into the metrics, not assumed from the request.
Built environments keep their editable Blender scene, GLB, images and source. Robot meshes keep their Menagerie licenses.

What leaves your workspaceper run

Item	Goes to
Camera frames, pose, instruction, history	the provider you selected
Scene files, evidence archives	stay in your workspace
Provider credentials	never in evidence or programs
Model-written programs	sandbox, no network

privacy

Whose data it is

Environments and runs are private to their owner. Model inputs, instructions and conversation history go to the provider you selected and nowhere else. Credentials never enter the evidence, and programs written by a model cannot reach the network or other episodes. You choose what to publish.