How far can code-only policies go?
We view the embodied world as an Embodied Turing Machine, whose tape is the robot and environment state and rules are the policy. What if a robot policy didn't need a VLM or VLA in its runtime control loop?
We want to explore the limit of a different paradigm: Can code-only-as-policy go surprisingly far in embodied intelligence? Where does it work, and where does it break?
- Data engine. Given offline time, coding agents iterate on a reusable code library. A working program then generates diverse trajectories efficiently and at scale for policy training.
- Stable, fast execution. Decisions are explicit code, and no model is queried at run time, so execution is stable, fast and cheap. This suits industrial settings where the state can be measured reliably.
- Generalization trade-off. Zero-shot generalization is difficult for code and generally calls for an agent in the loop. Code, however, is reusable: with offline development, it extends to new tasks and scenes.
One shared library. Cross-episode, cross-task.
The rules are code. The tape is the state.
One benchmark, as an example.
White box and controllable.
Every step reads and writes explicit state, never hidden in a model.
Backtrack and reattempt another way.
per control step. No model calls.
Easy to evolve new tasks.
COAP evolves fast, runs fast and works across episodes, so it generates trajectories efficiently for VLA.
Developed offline by coding agents.
A coding agent tries several approaches in parallel and keeps the one that works.
@article{hu2026embodied,
title = {Embodied Turing Machines: Stateful Code
for Robot Recursive Self-Improvement},
author = {Hu, Kairui and Hu, Siyuan and Hong, Fangzhou
and Chen, Zhaoxi and Liu, Ziwei},
journal = {arXiv preprint arXiv:2610.12369},
year = {2026}
}