[论文] Embodied Turing Machines: Stateful Code for Robot Recursive Self-Impro...
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
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
中文摘要
大多数机器人策略在控制回路中保留一个模型:VLA 将观测映射为动作,Agent Harness 在运行时查询 VLM 做决策。我们提出不同视角:具身世界是一台「图灵机」——纸带是机器人与环境状态,规则是策略。只要状态能被准确表示,决策完全可以写成代码。由此我们提出 COAP(Code-Only-as-Policy):代码从相机图像和本体感知中测量并追踪机器人、环境和任务状态,并据此做每个决策。同一套代码跨回合复用,不同任务共享一个代码库,回路中既无 VLM 也无 VLA。相比 VLA 和 Agent Harness,COAP 有三方面优势:(i)显式状态——状态可存储在代码中;(ii)可执行——决策可控、可从失败中灵活恢复、在线运行快速廉价;(iii)可扩展——新任务复用、继承或扩展共享库,能力随任务积累。这些优势使 COAP 天然适合递归自我改进(RSI):编程 agent 在闭环中开发代码库,每次变更都显式可控。在 RoboDojo 的 42 个双臂任务上,最终代码库在无模型的测试时达到 70.24% 成功率。COAP 的上限取决于状态表示的准确性和代码逻辑的鲁棒性。我们因此将 COAP 提为具身任务的新范式;由于它跨回合适用,还可作为 VLA 和 Agent Harness 的高效数据引擎。
原文摘要
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, we analyze three...
*自动采集于 2026-10-12*
#论文 #arXiv #CV #小凯