How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
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.
| 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).
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
| Doc | Path |
|---|---|
| RQ (framing note references this doc) | paper003_rq_v0.1.md |
| Method draft | paper003_description_v0.1.md |