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 — Method spec v1.0 (public)

Status: v1.0 · aligns with RQ v1.1 · pre-reg v2.0 in design


Pipeline

Nominal rollout (0 … onset−1) → record actions → mismatch @ S → replay → fork(response) → judge → profile

Same-state counterfactual evaluation: all branches share identical trajectory through mismatch onset S; only post-onset intervention differs. CF is method, not a learned recoverability estimator.


Mismatch onset S

Parameter Anchor (D0–D3) D1 control
Task Isaac-Reach-Dual-STAR-IK-Rel-Play-v0 same
Onset step 20 20
Target shift (P2) +0.06 m Y same
Occlusion (P3) L1 · visibility 0.35 · gain_scale v0.1 none
REPLAN delay 20 steps 20
Horizon 160 steps 160
Proposer Scripted 6D IK-Rel · body 13 same

Proxy contract: study1d_occlusion_proxy_v0.1.md.


Replay contract

  1. Fix seed · reset · run steps [0, onset) · store pre_actions.
  2. Capture frozen_xyz · apply shift → shifted_xyz.
  3. Per branch: reset(same seed) · replay pre_actions · diverge response from step onset.

Invariant: branch_replay_ok = true on all branches.

Forbidden AABB (EE proxy): center [0.45, 0, 0.15] · half-extent 0.04 m.


Response menu (D0)

Mode Behavior after S
CONTINUE Chase frozen target · occluded gain scale
REPLAN_d20 Frozen until t+20 · then shifted target
REOBSERVE 10-step hold · clear visibility · replan shifted
RESHAPE 20-step reshape proxy · clear · replan
HANDOVER Zero-action stub · early terminal

Runners: orbit_reach_study1a_counterfactual.py · orbit_reach_study1d_counterfactual.py.


Rule baselines (D2 · D3 · v2.0)

B2 — UQ-inspired binary: visibility & distance thresholds → HANDOVER_stub vs CONTINUE (rule proxy · not learned UQ).

B3 — Situation rule: occluded_shift → REOBSERVE path · shift_only → REPLAN_d20 · else CONTINUE.

Constants frozen in pre-reg v2.0. Desk spec: private baseline_spec_phase_c_v1.0.md.


Terminal judge

Primary — successful resolution:

Categories: successful_resolution · unsafe_failure · timeout_failure · safe_unresolved · handover_proxy.

HANDOVER excluded from primary endpoint · reported as burden proxy.


Claim tiers

Tier Label
A Scaffold / protocol
B Smoke n≤5 (001A–D)
C Pre-reg n=20 · D0–D3 executed · v2.0 complete 2026-07-24

Limitations

Scripted proposer · single sim task · occlusion proxy v0.1 · HANDOVER stub · B2 rule proxy · offline CF (not closed-loop deployment).