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 1 — research question (public lock)

Status: v1.1 locked · D0–D3 Tier C executed (2026-07-24) · working paper live
Scope: Stage 1 measurement · surgical reach proxy · EXP-SURG-001 unified with study1A–D
Program context: Non-Average World ruler @ fixed S — not WM extension · Study 002 separate (closed)


Central research question (v1.1)

EN:

At a fixed post-mismatch execution state S, do intervention classes yield separable recoverability profiles — and does a multi-mode menu outperform continue, a UQ-inspired binary rule, and a situation rule on resolution and burden metrics?

KR:

과업 관련 mismatch 직후 고정 상태 S에서, response class별 recoverability profile이 분리되는가 — multi-mode menu가 continue·UQ binary rule·situation rule보다 resolution/burden에서 나은가?

(v1.0 timing joint form retained as secondary object — delay confirmatory deferred; see RQ-T.)


Measured object

Intervention-conditioned recoverability profile (R_a(s,t)):

Counterfactual evaluation is method, not a claim of a new learned recoverability estimator.


Sub-questions (Paper 1 · v1.1 priority)

ID Sub-RQ Smoke (Tier B) Confirmatory (pre-reg v2.0)
RQ-M At fixed S, do response classes yield separable success profiles? (REPLAN ≫ CONTINUE) Direction yes — 001A · 001D ✅ D0 n=20 · REPLAN 19/20 vs CONTINUE 0/20
RQ-B Do multi-mode profiles beat B2 UQ rule and B3 situation rule? Not tested ✅ D2 · D3 · direction met
RQ-P Does occlusion change profiles vs shift-only @ matched 6 cm? 001D direction ✅ D1 · descriptive
RQ-C Do REOBSERVE / RESHAPE add viable paths beyond REPLAN? 001D smoke 4/5 Exploratory · D0 profile (17/20 · 18/20)
RQ-T Does delaying REPLAN materially reduce recoverability? Mostly flat · 001B/C Deferred — not confirmatory headline

Experimental arm map (EXP-SURG-001)

Arm Study Sub-RQ Tier Role
Mode @ 3 cm 001A RQ-M B Design · counterfactual grid
Delay grid 001B RQ-T B Defer confirmatory
Severity × delay 001C RQ-T · anchor B 6 cm · d20 lock
Occlusion + 5-mode smoke 001D RQ-M · RQ-C · RQ-P B Proxy validation
Primary cell D0 (Phase C) RQ-M · RQ-C C Executed 2026-07-22
No-occlusion control D1 RQ-P C ✅ 2026-07-24
B2 UQ-inspired binary D2 RQ-B C ✅ 2026-07-24
B3 situation rule D3 RQ-B C ✅ 2026-07-24

Study 002 (EXP-SURG-002): informative scenario generation — separate program · optional appendix only · not Paper 1 body.


VAP-separating form (preferred framing)

Given an execution-time mismatch, can counterfactual intervention profiles identify when replanning, re-observation, environment shaping, or human assistance is most likely to preserve successful task resolution — compared to continuing or a fixed situation-handler rule?

We evaluate profiles; we do not claim a new VAP-TAMP-style architecture.


Claim tiers (honest labeling)

Tier Label Evidence
A Scaffold / protocol Same-state CF pipeline · ORBIT Isaac · replay OK
B Smoke / direction 001A–D · n=5 · scripted IK-Rel
C Confirmatory D0–D3 Phase C n=20 · pre-reg v2.0 · 2026-07-24

Public stance today: Tier A + B + one Tier C cell (D0). Full publication-grade program after D1–D3.


Method (Stage 1)

Shared rollout → mismatch @ S → replay → fork(response [, baseline rule]) → terminal judge → profile table

Platform: Isaac Sim 4.1 · ORBIT Dual-STAR Reach · scripted 6D IK-Rel.

Anchor cell: +6 cm Y @ step 20 · occlusion L1 (visibility 0.35) · REPLAN_d20 · n=20.


Paper 1 contributions (scoped public)

# Contribution
1 Same-state counterfactual evaluation — fair intervention comparison at mismatch onset
2 Intervention-conditioned recoverability profiles — empirical slices 001A–D + D0 proper
3 (v2.0) Rule baseline comparison — UQ-inspired binary · situation rule @ same S

What Paper 1 is not



Version history

Version Date Note
v0.9 2026-07-16 Central Q + Stage 1 method
v1.0 2026-07-22 Sub-RQs · claim tiers · post-smoke lock
v1.1 2026-07-24 Unified EXP-SURG-001 arms · RQ-B/P promoted · RQ-T deferred · D0–D3 matrix · Study 002 boundary