Current robotic systems lean heavily on large neural network policies — either running Vision-Language-Action models end-to-end or querying VLMs at high frequency. This approach brings inference latency, high deployment costs, and unpredictable failures. Researchers from NTU Singapore propose a radical alternative: treat the physical environment as a Turing machine and execute robot control entirely through deterministic, stateful code — with zero neural network models at test time. Their Code-Only-as-Policy (COAP) system achieved 70.2% success across 42 complex bimanual manipulation tasks, proving that code can replace neural inference for embodied intelligence.
What Happened
The team from S-Lab at NTU Singapore introduces Embodied Turing Machines and Code-Only-as-Policy (COAP) in a new paper (arXiv:2610.12369). The core idea is to model the robot and its environment as a Turing machine: the tape encodes robot proprioception and visual states, while transition rules are executed by deterministic code. This code, drawn from a shared modular library, governs all decisions across episodes without running any neural network during deployment.
Crucially, the code-based approach enables Recursive Self-Improvement (RSI). Autonomous coding agents iteratively author, debug, and refactor robot skills in closed simulation loops. Because the policy is explicit code — not learned weights — improvements can be validated, versioned, and composed with full transparency. The system was tested on 42 complex bimanual manipulation tasks, achieving a 70.2% success rate with zero test-time models.
This result challenges the prevailing assumption that embodied agents must rely on neural networks for perception and action. COAP shows that a well-designed code library, combined with automated self-improvement, can match or exceed the performance of policy networks while eliminating their latency and cost overhead. The researchers position this as a paradigm shift for robotic deployment, especially in environments where predictability and low resource usage are critical.
My Take
This is one of the most refreshing ideas in embodied AI I’ve seen in a while. The field has become obsessed with scaling up VLMs and fine-tuning them for robotic tasks, but that path leads to bloated, brittle deployments. NTU’s approach takes a step back and asks a fundamental question: do we actually need neural networks to run the robot once the policy is learned? Their answer — no, we can compile the policy into deterministic code — is both elegant and practical.
For developers, this means lower cost and higher reliability. No more worrying about GPU availability at test time, no more non-deterministic inference errors. The recursive self-improvement loop is also a clever way to bootstrap complex behaviors without human hand-coding every edge case. I think we’ll see this concept extended to more domains, especially in industrial robotics where determinism is paramount. That said, the 70.2% success rate on 42 tasks leaves room for improvement — neural policies still have an edge on generalization. But the gap is closing, and the benefits in cost and predictability are immediate.
What to Watch
- How COAP handles open-ended perception tasks that require understanding novel visual scenes without any neural network — the current approach may need a lightweight vision encoder for state extraction, which is a potential weak point.
- Whether other labs adopt the Embodied Turing Machine formalism and build on the modular code library concept, potentially creating a standard framework for code-based robotics.
- The impact on edge robotics: If robots can run policies on microcontrollers with deterministic code, the cost of deploying sophisticated automation could drop dramatically.