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 — Lit positioning: execution-horizon framing vs internal RQ v0.1

Purpose: ground the "framing note" in paper003_rq_v0.1.md with actual papers, so the lab-intro framing and the internal RQ can be told apart on paper, not just in conversation. As of: 2026-07-31 — these are recent preprints, positioning may need a refresh before submission.


What the "execution horizon" literature actually studies

Paper Authors What it does
Dynamic Execution Horizon Prediction for Chunk-based Robot Policies (DEHP) Zhao, Bogdanovic, Sohal, Tao, Darvish, Aspuru-Guzik, Shkurti, Garg Learns how many actions of a frozen chunk policy to execute before requerying — trains a lightweight horizon head via online RL on top of a pretrained, unchanged policy
PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking — Adapts horizon per manipulation phase; discards unexecuted chunk suffix before replanning
Spatial Attention: Adapting Execution Horizons for Diffusion Policies via Observation Sensitivity — Horizon adaptation driven by observation sensitivity, still within a fixed diffusion policy
Deployment-Time Reliability of Learned Robot Policies Garg-affiliated Closest to "model adequacy" language, but at deployment-time monitoring level, not structural revision

Common thread: all of these hold the policy / model class fixed and ask how often should it be re-queried, and how far can it be trusted open-loop before the next observation? This is an L1-adjacent question (correction frequency within a fixed representation) — not an L3 question (whether the representation itself needs to change).


Where Paper 003 differs

Paper 003's RQ sits one level up: not "how often to replan" but "when repair within the current model class is insufficient, does changing the model class (adding a relation) unlock tasks that no amount of replanning frequency could reach?"

A chunk policy that requeries every single step is still bounded by what its representation can express. Paper 003 tests whether the binding constraint is representation adequacy, not query frequency — the axis these execution-horizon papers don't touch.

Execution-horizon literature:  WHEN to re-query a fixed model      (frequency)
Paper 002 / 003:               WHETHER the model class itself      (structure)
                                needs to change, and whether that
                                unlocks tasks frequency cannot reach

Not claiming


Open item


Doc Path
RQ (framing note references this doc) paper003_rq_v0.1.md
Method draft paper003_description_v0.1.md