After the Spill

How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.

View the Project on GitHub higuseonhye/research-os

Paper 003 — the two-body encounter, and what still blocks the endpoint

Correction, 2026-08-04. The tolerance argument below is wrong. It reasoned that a forked task family does not inherit its predecessor's 20 mm, so the figure was open pending a new scene. The new scene exists, already has a rigid object, and sets the same 20 mm independently — see the Branch B scene note. The tolerance is settled, not open. Everything else here stands.

Design note, 2026-08-04. CPU proxy only, not preregistered, not run in Isaac. Written after the probe sweep showed arm B landing 0.90 on the treatment condition.

The tension this resolves

Fixing the relation gate required the reference to withdraw: until the target is observed after the reference leaves, a struck target and one still sliding from an earlier strike are the same history. But the withdrawal is also what leaves the target stationary at the commitment, and a stationary target is one that zero-order predicts exactly.

Measured directly, with a single reference body:

Encounter Arm B Displacement at commit Gate fires Commit estD
burst (advance only) 0.55 16.4 mm 36 26 0.00
probe (with withdrawal) 0.85 1.7 mm 26 53 0.68

Hard but unidentifiable, or identifiable but easy. Sweeping the reference speed from 4 to 80 mm/step and the coupling gain to 0.9 moved neither row; a "probe once then pursue" schedule produced eligible cells in 8 of 40 seeds.

The design

Two bodies. A prober contacts the target early and leaves for good. A pusher arrives later and contacts during the dispense window.

At the commitment the target is stationary — the prober has gone, the pusher has not arrived. So zero-order predicts no motion, and so does constant velocity. Only a model that applies the observed relation to a second body predicts the displacement.

That is a stronger test than the single-body version, which arm D could satisfy by extrapolating one body's trajectory. Here the relation must generalise to an instance it was not learned on.

Measured on the CPU proxy, 0.5 mm observation noise, ~90 cells per row:

Coupling gain Arm B Arm C Arm D D* estD Displacement
0.50 0.63 0.33 0.91 1.00 0.74 15.6 mm
0.70 0.57 0.32 0.92 1.00 0.78 17.3 mm
0.90 0.58 0.27 0.92 1.00 0.77 17.1 mm

Arm D is clearly ahead, the coupling estimator fits in ~3 of 4 cells, and arm C is worse than zero-order — it extrapolates the stationary target's noise.

Eligibility, on arm-neutral grounds

A cell counts when the pusher will enter the interaction radius early enough in the dispense window for the contact to act. Two things are being excluded, and neither refers to any arm:

The existing motion_expected() predicate does neither correctly. Measured against ground truth it admits cells where the target does not move in 51% of cases under burst and 84% under probe — it uses the instantaneous closing rate as a proxy for future contact, which a schedule with reversals breaks.

Two findings that change preregistered choices

Reference speed does not grade difficulty

The draft grades the variant set T by reference speed. Measured, it does not grade anything:

Pusher speed (mm/step) 10 15 20 25 30 40
Arm B 0.69 0.68 0.67 0.71 0.69 0.84
Displacement (mm) 10.6 12.5 12.6 13.0 11.8 12.7

Flat across a fourfold range. A faster body penetrates further per step but spends fewer steps in contact, and the two offset. A dimension that does not grade cannot define a variant set, and this is a statement about the mechanism rather than about any arm's score.

Coupling gain grades weakly — displacement 10.5 → 17.2 mm from gain 0.3 to 0.85 — and is the better candidate, but see below.

Arm B never approaches zero, and the tolerance is why

Across every combination tried — both encounters, speeds from 4 to 80 mm/step, gains from 0.3 to 0.95, and eligibility margins from 1 to 5 steps — arm B lands between 0.55 and 0.90. It was never near zero.

The reason is structural. The coupling pushes the target away, which reduces penetration, which reduces the push: it is self-limiting, and one pass displaces the target by roughly 15 mm. The success tolerance is 20 mm. A displacement smaller than the tolerance means zero-order lands, whatever the arms do.

The CPU proxy that produced arm B in a 0.05–0.09 band did so at a 5 mm tolerance, on a scale the proxy chose arbitrarily. The 20 mm figure is inherited from Paper 001/002.

H1's capability-crossing endpoint is therefore not reachable at a 20 mm tolerance with this coupling. Either the displacement must be several times larger, or the tolerance must be smaller.

What this does not license

Choosing the tolerance so that arm B fails. That is selecting a parameter by which arm it favours, and it is the move the preregistration exists to prevent.

The legitimate route is already open and was recorded this morning for an independent reason: the scene reports rigid_objects: [], so real contact requires adding a body, which forks the task family. A forked family does not inherit its predecessor's tolerance — the 20 mm figure was established for a scene this no longer is. The new tolerance must come from the new scene's own geometry: the size of the object being placed and of the region it must land in.

Until that scene exists, the tolerance cannot be set, and until the tolerance is set it cannot be known whether the capability-crossing endpoint is reachable at all.

Status

Item State
Two-body encounter implemented in relation_dynamics and commitment_episode, 130 tests passing; not yet in the Isaac runner
motion_expected() known-wrong under reversing schedules; needs the eligibility rule above
Variant grading dimension speed shown non-functional; gain is the candidate
Placement tolerance blocking, and cannot be set before the Branch B scene
H1 reachability unknown, and downstream of the tolerance

Implemented 2026-08-04

The gate, the coupling estimator, normal_alignment and the projection now take either one body or several. A single-body history is stored as a one-body list, so nothing that passes one array changes behaviour — pinned by a test.

Three decisions the implementation forced:

Every statistic follows the nearest body. The interaction radius is smaller than the separation the encounter keeps between bodies, so only the closest can be in contact. Pooling all bodies would let a distant body's stillness dilute a real contact.

The projection follows the acting body, not the nearest one. After the prober leaves it can remain closer than the pusher for several steps while having no further effect; rolling the prediction forward with it would predict a contact that is over. The acting body is the one closing fastest, or one already inside the radius — observable, and independent of any arm's model.

A body that never pauses had to become predictable. The pattern estimator looked for a burst cycle and refused anything without a completed pause, so the pusher — which sits still, then closes at constant speed — was unpredictable and arm D could never act on it. It now falls back to constant motion, but only after an unbroken run of at least twice the horizon. That threshold is what stops the fallback reviving an older defect: a bursting body that has not paused yet also shows no completed cycle, and predicting continuous motion for it is exactly the error that once let a commitment land at step 7 of a 14-step cycle with arm D silently degraded into arm B. At 10-on/4-off with a 6-step horizon, twelve unbroken steps cannot occur.

Measured end to end through the real driver, 80 cells, 0.5 mm noise, gain 0.7: arm D 0.97, arm B 0.80, arm C 0.42, gate firing in every cell and arm D engaging in 0.89. Arm B is high here because the old eligibility predicate is still in place and admits stationary cells; the corrected predicate above put it at 0.57–0.63 in the standalone measurements.

A property worth stating rather than fixing: if the second body pushes without pause the target settles into a steady drift, a constant-velocity model explains it, and the gate declines — correctly, since that is arm C's case. The commitment has to precede the sustained push, which is what the eligibility window arranges. Pinned by a test.

Eligibility, implemented 2026-08-04

motion_expected() now takes the bodies' positions over the dispense window, supplied by the harness. That source matters: predicting the future instead would route eligibility through arm D's pattern estimator and make it depend on one arm's readiness, which is the coupling this design has had to undo three times. The Isaac runner computes the schedule, so it knows the future exactly.

A cell is admitted when the target is already moving fast enough to leave tolerance, or when a body occupies at least min_contact_steps of the window. Two steps is the default, set from the structure of the action rather than any arm's score: the placement lands at the end of the window, so a contact beginning on the final step has no time to move anything.

Measured against ground truth over 80 two-body episodes:

Eligibility rule Admitted Precision B C D Displacement
instantaneous proxy 915 0.37 0.80 0.42 0.97 3.7 mm
future, ≥1 contact step 757 0.45 0.77 0.27 0.99 7.5 mm
future, ≥2 contact steps 673 0.51 0.76 0.25 0.99 7.6 mm
future, ≥4 contact steps 514 0.57 0.75 0.17 0.98 8.7 mm

Precision is the fraction of admitted cells where the target really moves beyond tolerance. The old proxy was right about a third of the time; the corrected rule about half. The instantaneous fallback is kept for single-body callers and documented as the proxy it is.

Arm B barely moves — 0.80 to 0.76 — and that is the expected result, not a disappointment. Even with perfect knowledge of the future and a guaranteed two-step contact, the median displacement is 7.6 mm against a 20 mm tolerance. The residual imprecision is the tolerance, not the predicate. This is the same finding as above, arrived at from a different direction.

The runner, 2026-08-04

The second body needed no scene construction at all: the reference bodies are moving points, not rigid assets, so the runner computes their positions analytically and hands them to the driver. A second analytic trajectory costs nothing. --bodies 2 now runs the encounter under Isaac with injected coupling.

That is a validation of the machinery, not progress on real contact. Deliverable 2 stays unmet: the scene reports rigid_objects: [], so contact physics still requires adding a body.

While doing it, all of the geometry moved out of the runner into wm_expansion/encounter.py and under test. It had been inline in a file that cannot be imported without a GPU — the surface that produced eight defects, among them the fixed +x axis that turned ten seeds into one translation-invariant encounter. The runner is now 426 lines and decides almost nothing; 19 tests cover the schedules, the draw, and the body positions, including that adding a second body leaves the first's trajectory untouched.

First Isaac run, 2026-08-04

One cell on the reach task with injected coupling, --bodies 2. The encounter appeared exactly as designed, and both defects the CPU loop had caught were absent:

Steps Body 1 Body 2 Target
0–6 106 → 38 mm, strikes 230 mm moves 1.8, 8.4, 11.6 mm
7–16 withdraws to 113 mm, stops 219 mm completely still
17–27 113 mm 219 → 59 mm still
28–32 126 mm 48 → 43 mm, contacts moves 2.0, 6.5, 7.0, 4.9 mm

Closest approach 38.4 mm and 43.0 mm — both inside the 50 mm radius, so both bodies did their job. The gate fires at step 11, right after the post-departure observations accrue and long before the second body arrives, which is the identifiability argument working in the simulator rather than on paper.

Arms B and C also landed in this cell: the second body's contact displaced the target about 22 mm in total, and less than 20 mm of that fell inside the dispense window. One cell says nothing, and that quantity is exactly what real contact will decide rather than us.

What remains

  1. The placement tolerance, which is blocking and needs the Branch B scene.
  2. The Branch B scene itself — a rigid body in the environment, so that contact is simulated rather than injected. The orbit-surgical lift and handover task folders, which do have rigid objects, are deliberately deleted by the bootstrap script as "incompatible", with no record of why. Finding out is a two-minute check on the pod and is the cheapest next GPU action.

The tolerance moves every number here, so it should settle before any figure from (2) is treated as more than a smoke test.