RobotGym by Moving Atoms
Try this environment →
PUBLIC EXAMPLE · PRECISION ELECTRONICS

A precision task. An inspectable robot run.

The full simulated electronics testing laboratory with actual simulator robot

Editable laboratory, workbench, storage, instruments, guards and task fixture. Robot geometry comes from the simulator.

Precision has to exist in the physics

Small holes.
Real collision geometry.

Pick the populated board, align its mounting holes, engage the round and relieved locating pins, release and inspect. Then retrieve vertically and place it in OUTPUT.

85 × 56 mmManufacturer PCB outline
2.7 mmMounting-hole diameter
0.35 mmNominal round-pin radial clearance
96 piecesRigid compound collision geometry

The original CAD provides populated components, connectors and open mounting holes. Convex collision pieces preserve those openings instead of filling them with a hull.

Actual robot above the workbench
Close-up of the populated board seated on the fixture
Board placed in the output nest and robot withdrawn
GPT-6 Astra · actual recorded executionDownload video ↓
27.4 sSimulation time
5Model calls
0.14 mmMaximum seated hole-centre error
8.65 NPeak board–fixture contact

The policy receives an explicit commissioned setup with nominal coordinates and grasp settings. It writes bounded robot programs and inspects camera checkpoints. Physics pauses while it thinks. This is a guided nominal-layout evaluation, not a vision-only pose-estimation benchmark.

The main view renders recorded physical poses in Blender; all three camera panels replay the same frozen simulator states. Motion is neither generated video nor hand animation. The original run is preserved; measured results and commands are available below.

Measure the task, including its failures.

We declared the acceptance limits before the native evaluations. Independent command replay matched every recorded state exactly, then checked contacts at 500 Hz. Earlier failures remain visible.

EvaluationVerdictValid released seatingPeak board–fixturePeak robot–sceneOutput centre error
Attempt 1Scene 311ea8e6Rejected0.00 s39.13 N38.67 N301.48 mm
Attempt 2Scene 311ea8e6Rejected0.00 s32.15 N8.37 N1.02 mm
Attempt 3Scene 311ea8e6Passed5.56 s8.49 N0.00 N0.80 mm
Attempt 4Scene 31879de4Passed5.56 s8.49 N0.00 N0.80 mm
Attempt 5Scene 5377e138Rejected0.00 s22.03 N0.00 N0.93 mm
Attempt 6Scene a75c3615Passed5.56 s8.65 N0.00 N0.65 mm
549States matched exactly
5.56 sValid released fixture dwell
0.00 NRobot contact with static geometry
0.65 mmFinal output placement error

The first evaluation stopped after a failed retrieval. The second completed the motions, but the audit rejected a tilted seat and excess force. A corrected procedure passed. Later scene revisions restored display details and replaced approximate component envelopes with CAD bounds, exposing a turning failure. The final passing setup uses an aligned input nest. These are development iterations across changing scenes and instructions, not a measured success rate.

Acceptance: both locating pins contained within the holes, height ±0.20 mm, tilt ≤0.5°, released dwell ≥1 s, board–fixture contact ≤15 N, robot–scene contact <1 N, and output centre error ≤1 mm. Force values are evaluator-only simulated contact-normal sums, not physical sensor readings or certified safety limits.

Make this environment yours. Copy the scene, robot and camera setup, and commissioned task into your workspace. Change the task or policy and run your own experiment. Sign-in is required to save a copy; normal compute credits apply when you run.