Most robot policies today keep a learned model in the control loop: a vision-language-action (VLA) model maps observations to actions, while an agent harness such as Agent-as-Policy or Harness VLA queries a vision-language model (VLM) at run time to make decisions. A new paper on Hugging Face proposes a different framing: the embodied world is an Embodied Turing Machine, where the tape is the robot and environment state and the rules are the policy.
The authors argue that if this state can be represented accurately, decision making can be written entirely in code. Their approach, Code-Only-as-Policy (COAP), uses code to measure and track the robot, environment, and task state from camera images and proprioception, and makes every decision from that state. The same code applies across episodes, and different tasks share one library without a VLM or VLA in the loop.
The paper analyzes three advantages of COAP over VLAs and agent harnesses: explicit state, where state is stored in code; execution, where code makes decision making controllable, allows flexible failure recovery, and runs fast and cheaply online; and extensibility, where new tasks reuse, inherit, or extend the shared library so capabilities accumulate over tasks.
These properties, the authors suggest, make COAP a suitable medium for recursive self-improvement (RSI): coding agents develop the library in a closed loop, and each change is explicit and controllable. On RoboDojo's 42 bimanual tasks, the resulting library reaches a 70.24% success rate without a model at test time.
The authors note that COAP's upper bound depends on how accurately state is represented for decision making and how robust the code logic is. They position COAP as a new paradigm for embodied tasks and, because it applies across episodes, as a potential efficient data engine for VLAs and agent harnesses.
The paper states that code was developed on development episodes, benchmark test episodes were not used for development, and oracle values are not used in code. The authors emphasize this is not an attempt to achieve state-of-the-art on RoboDojo or to hack a benchmark, and that it should not be misread as a rival to world-action models, VLAs, or agentic methods; rather, it is meant to show how far code-only RSI can go and to offer the community a way to generate trajectories efficiently in a simulator.