RobotyWhat did running Reach on an iPhone make possible?
The phone executed this authored simulation itself. Comparing it with a Mac reference checked software agreement while leaving physical accuracy and other tasks open.
Covers October 9, 2026. Reconstructed later from available records.
On October 9, the iPhone ran the physics for Reach, our simulated arm-to-target task. It no longer needed a Mac to execute this particular simulation.
That changes a practical dependency. The phone could do more than display a result produced elsewhere. It could execute the supported authored task locally and retain the attempt.
The device was an iPhone Air running iOS 27.0.1 with MuJoCo 3.14.0 and a precompiled Reach model. Three local attempts completed successfully. Each advanced through 280 control steps over 0.56 simulated seconds and retained 281 samples, including the initial and terminal states. Simulation time does not tell us how long someone waited. These records did not measure execution wall time.
Check that the two implementations agree
The first phone attempt was compared with a separately executed Mac reference at seven checkpoints. Both used the same MuJoCo engine, model and authored task through different Swift/C and Python implementations.
All seven checks passed. The allowed differences were 0.0001 radians for joint coordinates, 0.0001 metres for tool position and one billionth of a simulated second for dwell. The observed joint and position differences were about 5.55e-17 in their respective units, with no dwell difference. These quantities do not combine into one accuracy score.
That checks agreement between software executions at the measured points. It does not independently validate physical accuracy, every intervening step, or every possible target. The earlier task-evaluation work explains why moving and satisfying the goal are separate questions.
A short run leaves other questions open
The second and third attempts finished before the first post-Start accessibility snapshot. No live Stop control was observed or tapped. Backgrounding happened after a terminal result. A completed run cannot demonstrate cancellation.
The useful result is specific: this authored simulation executed on an actual iPhone. Moving one supported task off the Mac removed that dependency for this task. It did not establish that every Roboty feature could run without one.