Embodied Turing Machines: NTU Singapore Proposes Code-Only-as-Policy (COAP) for Robots, Achieving 70.2% Success with Zero Test-Time Models

Background and the Problem

Embodied robotics is increasingly dominated by neural network policies: end-to-end Vision-Language-Action (VLA) models or agent harnesses querying cloud VLMs at high frequencies. This paradigm introduces severe engineering bottlenecks: high inference latency incompatible with 20Hz+ control loops, unsustainable token costs, non-deterministic execution failures, and lack of interpretability when hardware fails.

Architecture and How It Works

Researchers from NTU Singapore S-Lab propose Embodied Turing Machines and Code-Only-as-Policy (COAP, arXiv:2610.12369, site: anonymous-report-421.github.io/public-website/):

1. The World as an Embodied Turing Machine: Formulates physical environments as Turing machines where the tape stores environment and robot states, while the transition rules are executed by deterministic code. Vision extractors map pixels into state variables, while control and failure-recovery logic are written purely in stateful code.
2. Three Core Architectural Advantages:

• Explicit State: Internal variables are stored in transparent code structures.
• Deterministic Execution: Runs sub-millisecond on embedded CPUs with native try-catch exception handling and zero per-step cloud costs.
• Modular Extensibility: Skills are modular software libraries that can be inherited, imported, and shared across heterogeneous tasks.

1. Coding Agent Closed-Loop Recursive Self-Improvement (RSI): Because policies are pure code, Coding Agents inspect execution logs, edit Python routines, and continuously refactor robot skills within simulation harnesses.

Benchmarks and Measured Results

Benchmarked across 42 demanding bimanual manipulation tasks in RoboDojo:

1. 70.24% Success Rate with Zero Runtime Models: Synthesized code libraries achieve a 70.24% resolve rate across 42 complex bimanual tasks without invoking any neural network during deployment.
2. Over 99% Latency Reduction: Drops decision latency from 1,000ms+ down to under 1ms, fulfilling strict industrial control cycle requirements.
3. Synthetic Data Engine Capability: Deterministic COAP trajectories serve as high-fidelity data engines, generating millions of verified demonstrations to train downstream VLA baselines.

Getting Started for Developers

The COAP report establishes a practical blueprint for robotics and automation engineers. Instead of treating every task as an end-to-end black box, teams should decouple state estimation from policy logic. Use robust perception to extract geometric poses, express manipulation routines in modular Python libraries, and deploy coding agents (such as Claude Code) to autonomously maintain and debug robot codebases inside CI/CD test sandboxes.