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 — real-contact coupling, design v0.1

Status: design only, 2026-08-03. Nothing here has run. Addresses: deliverable 2 of the draft preregistration, the one the pilot could not meet.


What is wrong with the current coupling

The pilot moves the target by calling coupling_displacement, a function in this repository, and arm D predicts by inverting that same function. The loop is closed on itself. Two consequences:

  1. No physical jitter exists to measure. Deliverable 2 asks for observation noise and timing irregularity under real physics. The current setup generates neither, so the noise levels the estimator was characterised against are assumptions rather than measurements.
  2. Arm D's model is exactly correct. It fits gain × (radius − d) to data generated by gain × (radius − d). Recovery is near-exact on clean data, and that says nothing about whether a prepared relational model helps when the world does not match it.

The second is the more serious. What the paper claims is that a prepared relation operator earns its place — not that a model can invert itself.


What real contact changes

Under PhysX the target's response involves mass, friction, restitution and contact geometry. It is not linear in separation. So arm D's linear module becomes misspecified, which is the interesting case:

Injected coupling (now) Real contact (proposed)
Arm D's model exactly right approximately right at best
Physical jitter none emerges from contact
What a win means the function is invertible a prepared linear relation is useful despite being wrong
What a loss means implementation bug the prepared operator does not transfer

Arm D may well get worse, and that is a legitimate outcome. It should be declared now, before the run, exactly as the commit-policy direction was: moving to real contact is expected to reduce arm D's accuracy, because its model stops being exact. The reason for doing it anyway is that the current number is not a measurement of anything the paper claims.


Two implementation branches

Which applies depends on what the task's scene already contains, and that cannot be determined without a GPU. The pilot runner now records a scene_inventory (articulations, rigid objects, sensors) in every record, so one cheap run answers it rather than guesswork committing several hundred lines of scene code.

Branch A — a rigid body already exists

If Isaac-Reach-Dual-STAR-IK-Rel-Play-v0 ships a manipulable object, reuse it:

Smallest change, and it keeps the task recognisably the one Paper 001/002 used.

Branch B — no usable rigid body

A custom scene configuration is needed, adding two rigid objects to the existing environment cfg. Larger, and it forks the task family away from the one prior papers used, which weakens the inherited tolerance argument — the 20 mm criterion was borrowed on the grounds that it is this task's established threshold.

If Branch B is required, that inheritance has to be re-argued or the tolerance re-derived, and that should be settled before the run rather than after.


What stays fixed either way

Decisions already locked do not reopen because the physics changed:


What must be re-derived

Each of these is a preregistration parameter. They are re-derived from the pilot, never from confirmatory data.


Order

  1. One pilot run on the current build → read scene_inventory → pick a branch.
  2. Implement that branch; smoke a single cell and read the record by eye.
  3. Confirm the target actually moves on contact and that the arms see it.
  4. Re-derive the parameters above.
  5. Only then freeze the preregistration.

Doc Path
Draft preregistration paper003_prereg_draft_v0.1.md
Pilot runbook paper003_pilot_runbook_v0.1.md
Pilot results, including why deliverable 2 is unmet isaac_relation_pilot_v0.1/RESULTS.md