$ cat wiki/papers/2026/2610.12369-embodied-turing-machines.md
Embodied Turing Machines — code-only robot policies
TL;DR
Code-Only-as-Policy (COAP) represents robot and environment state explicitly and makes decisions in reusable code, without a VLM or VLA in the execution loop. The authors report 70.24% success on RoboDojo. (source)
Authors & Org
Kairui Hu, Siyuan Hu, Fangzhou Hong, Zhaoxi Chen and Ziwei Liu. Submitted 2026-10-08. The captured abstract page does not state affiliations. (source)
Method
Code measures and tracks state from camera images and proprioception. A shared library supports multiple tasks and episodes; coding agents develop it through a closed loop in which changes remain explicit. The authors describe explicit state, controllable execution and library reuse as the basis for recursive improvement. (source)
Results
The resulting library achieves 70.24% success across 42 bimanual RoboDojo tasks, with no model at test time. The abstract provides no per-task table, baseline score, or end-to-end development cost. (source)
Significance
This connects Agents (LLM Agents) with explicit state in World Models. It complements Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks, which compiles revisable world models into executable code: both place useful behavior outside a changing LLM, but evaluate different tasks and cannot be ranked together. (source) (Memento 3)
Open Questions
The authors identify state-estimation accuracy and code robustness as the limits. Whether the library generalizes beyond the tested task collection, and how much agent development costs, remain unanswered by the abstract. Full paper and physical deployment were not reviewed. (source)