How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Finding, 2026-08-04. Measured in Isaac on the lift scene; the fix is not implemented, and an attempt to patch it is recorded here as a warning.
In Isaac-Lift-Block-PSM-IK-Rel-Play-v0, with the encounter's script speed
lowered until the arm could track it:
min observed separation 10.1 mm
min commanded separation 10.2 mm
interaction radius 12.0 mm
The arm follows the script almost exactly — 10.1 against 10.2. The script itself never asks it to come closer than 10.2 mm, so the closest approach barely enters the radius, and no cell produced a gate firing or an engaged relational arm.
draw_geometry places the first body at 2.2 to 2.8 interaction radii and
advances it for probe_advance steps. Those proportions were drawn for a 50 mm
radius:
| radius 50 mm | radius 12 mm | |
|---|---|---|
| Approach drawn | 110–140 mm | 26–34 mm |
| Advance covers | 7 × 15 = 105 mm | 7 × 3 = 21 mm |
| Ends at | 5–35 mm — inside | ~9 mm short |
| Plus lateral offset | still inside | closest pass ≈ 10.8 mm — outside |
Two constraints were already enforced — the withdrawal must clear the radius, and a body must not travel further per step than the radius. The advance covering the approach was not, and it is the one that broke.
Drawing the approach as a fraction of what the advance can cover fixes the first body and breaks the second: a deeper strike displaces the target further, and the second body's line was fixed at draw time against the target's original position, so it misses. Four of twenty seeds lost contact entirely.
Aiming the second body at the object continuously fixes that and breaks arm D: a body that homes on a moving target curves, and the pattern estimator assumes a straight line at constant speed, so the relational arm fell below zero-order (0.45 against 0.60). A homing pusher is also the wrong phenomenon — it has intent, and this is meant to be a body that happens to strike.
Aiming once at launch and then travelling straight fixes the curve and leaves
arm C dominant (0.75 against arm D's 0.15), which suggests the deeper strike
leaves the target moving steadily enough for constant velocity to explain it —
the slide regime again, arrived at from a different direction.
Each patch fixed one thing and broke another, and the property tests flapped with them. That is the signature of a design that needs one coherent pass rather than a sequence of local repairs, so the branch was reverted and the suite left green at 206 tests.
Constraints 2 and 3 are in tension, and that tension is the actual design problem. Two candidates, neither tried:
| Item | State |
|---|---|
| The lift scene runs cells end to end | yes, contact=physical |
| The arm tracks the script | yes, within ~5 mm structurally |
| The encounter reaches the object | no — the finding above |
| Fix | not implemented; needs one pass, not patches |
| Any arm-level number from this scene | not yet meaningful |