Summary
This paper introduces the Embodied Turing Machine perspective, where the robot and environment state act as the tape and the policy as the rules, enabling decision-making to be written entirely in code. The authors propose COAP (Code-Only-as-Policy): code measures and tracks robot, environment, and task state from camera images and proprioception, making every decision without a VLM or VLA in the control loop. The same code applies across episodes, and tasks share a single code library. Compared with VLAs and Agent Harnesses, COAP offers explicit, storable state; executable, controllable decisions with fast, cheap online inference and flexible failure recovery; and scalable reuse, inheritance, and extension of a shared task library. These properties make COAP naturally suited to recursive self-improvement (RSI), where a coding agent develops the codebase in closed loop with explicitly controllable changes. On 42 bimanual tasks in RoboDojo, the final codebase achieves 70.24% success rate in a model-free test-time setting. COAP's ceiling depends on state-representation accuracy and code logic robustness. The authors position COAP as a new paradigm for embodied tasks and, due to its cross-episode applicability, as an efficient data engine for VLAs and Agent Harnesses.
论文概要
研究领域: CV
作者: Kairui Hu, Siyuan Hu, Fangzhou Hong, Zhaoxi Chen, Ziwei Liu
发布时间: 2026-10-08
arXiv: 2610.12369
Abstract (English)
Most robot policies keep a model in the control loop: a VLA maps observations to actions, and an Agent Harness, such as Agent-as-Policy or Harness VLA queries a VLM for decision making at run time. We propose a different view: the embodied world is an Embodied Turing Machine, whose tape is the robot and environment state and rules are the policy. If this state can be represented accurately, the decision making can be written entirely in code.
We therefore propose Code-Only-as-Policy (COAP): code measures and tracks the robot, environment, and task state from camera images and proprioception, and makes every decision from it. The same code applies across episodes, and different tasks share one library without a VLM or VLA in the loop.
Compared with VLAs and Agent Harnesses, COAP has three advantages:
- (i) Explicit state — state can be stored directly in code.
- (ii) Executable — decisions are controllable, flexibly recoverable from failures, and fast and cheap to run online.
- (iii) Scalable — new tasks reuse, inherit from, or extend the shared library, so capabilities accumulate with tasks.
These advantages make COAP naturally suited to recursive self-improvement (RSI): a coding agent develops the codebase in a closed loop, with every change explicitly controllable. On 42 bimanual tasks in RoboDojo, the final codebase achieves a 70.24% success rate in a model-free test-time setting.
COAP's ceiling depends on the accuracy of state representation and the robustness of code logic. We therefore position COAP as a new paradigm for embodied tasks; because it applies across episodes, it can also serve as an efficient data engine for VLAs and Agent Harnesses.
---
Paper: arXiv:2610.12369
*Auto-collected on 2026-10-12*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178635896