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 — a struck object slides, and that is arm C's case

Reversed the same day by measurement. The gate was run on real contact traces and fires on every trace containing a strike, with cv_gain negative throughout — a constant-velocity model is worse than zero-order on this motion. See real_contact_gate_v0.1.

The reasoning below is kept because the toy model and the retention statistic were built here, and because the failure is instructive: retention describes one coast in isolation, while the gate reads the whole episode. The proxy was not the gate, and should not have been read as it.

Finding, 2026-08-04. CPU, toy contact physics, not Isaac. The most consequential result of the design work so far, and it threatens the paper's premise rather than supporting it.

What was run

The cell loop now takes the target's pose from a World rather than computing it. Two implementations share every other line — the gate, the arms, the eligibility screen, the scoring:

ContactWorld was driven by a toy rigid body with momentum and friction: an approaching body imparts an impulse along the contact normal, and the object then coasts, decelerating by a friction coefficient. Twenty seeds per setting.

The result

Friction B C D D* Gate estD Displacement
μ = 0.35 0.85 0.25 0.85 1.00 0.00 0.00 16.8 mm
μ = 0.15 (slippery) 0.15 0.30 0.15 1.00 0.00 0.00 39.6 mm
μ = 0.60 (stops fast) 0.80 0.20 0.80 1.00 0.31 0.05 6.5 mm

Two things at once, and they pull in opposite directions.

Real contact makes the task hard, which is what was wanted. At μ = 0.15 the object is displaced 39.6 mm against the 20 mm tolerance and arm B falls to 0.15 — the near-zero band the injected coupling could never produce at any speed or gain.

And the gate never fires. estD is 0.00, so arm D falls back to arm B in every cell and the two are identical by construction. The operator the paper is about never acts.

Why, and why it was predicted

A struck object slides. Motion therefore continues while the bodies are far away, proximity_contrast collapses, and a constant-velocity model explains a good part of what happens. That is precisely the slide control — the condition added specifically because it "is what real rigid-body contact produces: struck objects slide" — and the gate declines it correctly.

The gate is not malfunctioning. It is reporting that under real contact the residual is substantially constant-velocity, which is Paper 002's operator.

The injected coupling never showed this because it is effectively μ = 1.0: the target stops the instant contact ends. The whole pilot series has been run in a regime that does not exist in a physics engine.

The question this reduces to

How fast does a struck block stop on the lift task's table?

The μ = 0.60 row is the clue: when the object stops within a step or two, the gate recovers to 0.31 and the encounter starts to look like the injected case. Below that it does not.

This is now a measurement, not a design argument. It is answered by striking the object in Isaac-Lift-Block-PSM-IK-Rel-Play-v0 and recording how many steps its velocity takes to reach zero, against the 6-step dispense window.

What follows either way

If the object stops quickly — one or two steps — the relation is proximity-conditioned in the sense the gate requires, and the design stands with real contact substituted for the formula.

If it slides — several steps — then the residual genuinely is part relation and part constant velocity, and the honest options are:

  1. Change the endpoint. The claim becomes about a composite residual that neither operator alone handles, which is a weaker but true statement.
  2. Change the task so the object cannot slide — constrained, or struck into a detent. That is a design change requiring an arm-neutral justification, and "so that arm D can act" is not one.
  3. Report that the relational operator does not survive real contact in this task family. A negative result, and a publishable one given the operator is Paper 002's own and the comparison is prepared.

None of these may be chosen by seeing which arm they favour. The measurement comes first.

How it is measured

wm_expansion/stopping.py takes a pose trace and the separation to the nearest body at each step, and reports when contact last occurred, how many steps the object then coasted, the per-step speed retention, and how far it travelled afterwards. All of it is under test, including the ways it could be quietly wrong: a trace with no strike, a strike that moved nothing, and contact running to the end of the recording are each refused rather than reported as a stopping time of zero.

It also emits a one-line gate_outlook, deliberately coarse and deliberately not a pass/fail, so a run states its own implication instead of leaving a number to be interpreted later — when the interpretation could be chosen.

scripts/orbit_lift_stopping_probe.py is the Isaac side: drive the end effector into the block, retreat, hold still, record poses. It has never been run.

/workspace/IsaacLab/isaaclab.sh -p scripts/orbit_lift_stopping_probe.py     --headless --seed 300 --out-dir results/paper003_stopping

The result is written to a file as well as printed, because Isaac's app swallows stdout — which cost a diagnostic earlier today.

Status

Item State
World abstraction implemented, 179 tests, injected path unchanged
ContactWorld implemented and exercised against toy physics
Isaac contact runner not started — needs the lift scene's names and a real pusher
Stopping-time measurement built: wm_expansion/stopping.py (12 tests) and scripts/orbit_lift_stopping_probe.py
The measurement itself the next GPU action, and everything waits on it
H1 reachability in doubt, for a reason that is now specific and testable