How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Public brief for Paper 002 Introduction · industry landscape · not a market report or vendor map
generalist policy / VLA / large behavior models
↓
world foundation models + synthetic data + simulation
↓
generation + reasoning + action simulation + modular experts + deployment flywheel
↓
(next) deployment failure → diagnosis → safe adaptation → selective reconstruction
Paper 002 sits in the last layer — not competing on foundation model scale.
| Actor | Role | Open gap Paper 002 addresses |
|---|---|---|
| NVIDIA (Cosmos · Isaac · GR00T) | Platform · sim · synthetic data · evaluation stack | When current model fails · repair vs expand · nominal safety protocol |
| Skild AI | Omni-bodied foundation brain · data flywheel | Which failure types require which reconstruction — not only more data |
| FieldAI | Deployment autonomy · unknown industrial sites | Failure-conditioned adequacy monitoring · selective autonomy reconstruction |
| Toyota RI | Large behavior models · rigorous evaluation | Failure taxonomy · inadequacy test · regression-safe adaptation |
| Physical AI / WAM | Policy scaling · world-action integration | Adequacy monitor above prediction/action layer |
| Surgical robotics (Medtronic · CMR · ecosystem) | Sim · validation · reasoning stack | Exception typing → recalibrate / replan / add mode / ask surgeon |
We evaluate failure-conditioned model adequacy and repair-vs-expansion decisions for deployed embodied stacks — not a new foundation world model.
Clinical long-term: tissue dynamics mismatch · unrepresented anatomy · observation reliability · external human action · tool–tissue relation change → different intervention classes (recalibrate · replan · add dynamics mode · change autonomy · ask surgeon).