Solutions · Research and benchmarks

Reproducible trials with declared budgets

Frozen scenes, seeds and limits. Independent grading. Evidence archives that reviewers can open. RobotGym gives a lab a bench that resets itself and remembers everything.

Start a project Read the evidence files

Same scene, same seed, same budget. If the result moved, something you can name moved it.

metrics.jsonfields from the current release · kitchen values

{
  "success": true,
  "termination": "model_finish",
  "control_steps": 329,
  "sim_seconds": 16.45,
  "wall_seconds": 262.02,
  "grade": { "lift_m": 0.082, "required_lift_m": 0.080,
             "left_pad_contact": true, "right_pad_contact": true },
  "limits": { … as declared before the run },
  "seed": 7
}

reproducibility

Frozen before the first step

The manifest freezes the robot, sensors, controller, scene version, budgets and source hashes before the first control step. Scene versions are content-addressed. Policy versions are written into the metrics as returned by the policy, not assumed from the request. The harness commit is pinned in the evidence.

Scene version and scene hash in every run. The kitchen run records 0a7f2919 and 1d824047.
Source photographs carry author, license and SHA-256.
Saved programs replay with identical actions on the same scene and seed.

Recorded trialspublic evidence

Recorded trial	Controller	Steps	Calls	Simulated	Result
Kitchen mug lift rebuilt from one photograph	Atom Code · GPT-6 Astra	329	10	16.45 s	success · lift 8.2 cm
Bowl onto plate LIBERO spatial · seed 7 · earlier release	π0.5 · hosted	81	17	4.05 s	success

benchmarks

Starters built from public benchmarks

The office desk starter is a reconstruction from RoboArena with an original and a shifted layout. An earlier release hosted LIBERO scenes, and one recorded π0.5 demonstration from it remains public. Two reconstructions from Creative Commons photographs, the kitchen and the electronics workbench, are public with their evidence.

Office desk from RoboArena, nominal and offset layouts.
LIBERO spatial scene with a recorded π0.5 trial, from an earlier release.
Kitchen and workbench reconstructions under their photographs’ licenses.

Two identical tabletop scenes side by side with objects shifted in the second — Nominal and shifted layoutsillustration

compare

Attempts side by side

History keeps every run on a scene, whoever ran it and however it ended. Termination reasons are separate fields from success. Today the harness drives the twin with GPT-6 Astra, Claude Fable 5 or Claude Fable 5.1 as the policy, through the same primitives, controller and budgets, so a difference between attempts belongs to the policy or the scene.

Spectator view, policy frames, commands, notes and verdict for every attempt.
Both clocks recorded. Physics pauses while a policy thinks.
Nothing is pruned to flatter a result.

Public filesmovingatoms.ai/showcase

publish

Evidence you can cite

Everything we say about a run traces to a file you can open. The kitchen evidence, the program the policy wrote and the earlier π0.5 demonstration are public. Cite runs by scene hash, seed and policy version. Students and labs can start on welcome credits.