Code-Only-as-Policy hits 70% on 42 bimanual RoboDojo tasks with no model in loop
Kairui Hu · hf · 2026-10-09
The author proposes viewing the embodied world as an Embodied Turing Machine: robot and environment state are the tape, policy is the rules—if state can be accurately represented, decision-making can be written entirely in code, dubbed Code-Only-as-Policy (COAP).
- Code measures and tracks robot, environment, and task state from camera images and proprioception, making every decision from it; the same code applies across episodes, with tasks sharing one library and no VLM or VLA in the control loop.
- Three advantages over VLAs/Agent Harnesses: explicit state stored in code; controllable execution with flexible failure recovery that is fast and cheap online; extensibility so new tasks reuse, inherit, or extend the shared library, enabling capability accumulation and recursive self-improvement via coding agents in a closed loop. COAP can also serve as an efficient data engine for VLAs.
- On RoboDojo's 42 bimanual tasks, the library reaches 70.24% success rate with no model at test time. The ceiling depends on state representation accuracy and code logic robustness.
More from Embodied
- China Begins Deploying Fleets of Self-Driving Delivery Robots for Overnight Operation — Polymarket · 2026-10-09
- Robotics isn't scaling because researchers haven't learned scaling like LLM folks have — eigenron · 2026-10-09
- Dev Ports Muse Linux Gadget SDK to .NET, Gets Edge Running on Arduino — unixterminal · 2026-10-09
- Photo to AI strokes to AxiDraw paper portraits: lessons from an AI Tinkerers Tokyo demo — Stefania_druga · 2026-10-09
- AssemblyAI turns nostalgic Furby toys into voice agents at SF Tech Week — AssemblyAI · 2026-10-09
- KAIST's ME-World tackles multi-agent egocentric world modeling with joint denoising — kaist-ai · 2026-10-09