Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement

Embodied Turing Machines. Stateful Code for Robot Recursive Self-Improvement. By Kairui Hu and four others, from Nanyang Technological University and Ropedia. Published on the eighth of October.

Most robot policies keep a model inside the control loop. A VLA maps observations straight to actions. An Agent Harness, such as Agent-as-Policy or Harness VLA, queries a VLM to make decisions at run time. Either way, every decision passes through a model while the robot is acting, and that model is the thing being trusted at every step. The question the paper raises is whether the model needs to be there at all, once an episode is under way.

The authors propose a different view. Treat the embodied world as an Embodied Turing Machine. Its tape is the state of the robot and its environment, and its rules are the policy. If that state can be represented accurately, then decision making can be written entirely in code. That is the proposal, named Code-Only-as-Policy, or COAP. Code measures and tracks the state of the robot, the environment and the task, working from camera images and proprioception, and every decision is made from that tracked state. The analogy carries the whole argument. In it, the tape and the rules are all there is. If the tape, here the state of the world, is read well enough, then the rules can be ordinary program logic, written once and run again.

The same code applies across episodes, and different tasks share one library, with no VLM or VLA in the loop. Against VLAs and Agent Harnesses, the authors analyse three advantages. First, explicit state: the state can be stored in code. Second, execution: code makes decision making controllable, recovers from failures flexibly, and runs fast and cheaply online. Third, extensibility: a new task can reuse, inherit or extend the shared library, so capabilities accumulate as tasks are added.

Those properties are what make COAP suited to recursive self-improvement, or RSI. Coding agents develop the library in a closed loop, and because the policy is code, each change they make is explicit and controllable. The library that grows this way is the robot's policy, edited directly rather than retrained.

The test bed is RoboDojo, with forty-two bimanual tasks, each needing both arms. The library that results from the closed loop reaches a success rate of seventy point two four percent across them. That figure comes without any model running at test time: every decision behind it was made by code.

The authors are direct about where the ceiling lies. How far COAP can go depends on how accurately the state is represented for decision making, and on how robust the code logic is. Those two factors set its upper bound. Put another way, the paradigm moves the hard problem rather than removing it: from learning a mapping to actions, to measuring the world accurately and writing logic that holds up.

COAP is offered as a new paradigm for embodied tasks rather than a single system. It also turns back toward the models it removes from the loop: because the code applies across episodes, the authors suggest it can serve as an efficient data engine for VLAs and Agent Harnesses. The robot's competence is moved out of weights and into a library that agents can read, extend and improve, and the measure of how well that works is how faithfully the code sees the world it acts on.