One shared library. Cross-episode, cross-task.
FastNo model at runtime.
StableDecisions are code, not samples. Every decision is explicit.
Cross-episodeOne program for every episode. Fixes carry over.
ReusableShared library. Easy to extend, reuse and inherit.
Data engineEfficient trajectory generation for VLA, Agent-as-Policy and Harness VLA.
The rules are code. The tape is the state.
cover_blocks.py· pseudocoderunning
Recorded statestep 0
head camerastep 000
No model in the online execution loop. Fast and cheap for data generation.
How far can Code-Only-as-Policy go?
success rate · RoboDojo, 42 bimanual tasks · 75.45 progress score
White box and controllable.
Explicit state
Every step reads and writes explicit state, never hidden in a model.
Controllable
76%failures caught
Backtrack and reattempt another way.
Cheap
0.3 ms
per control step. No model calls.
Extendable
Easy to evolve new tasks.
Data engine
COAP evolves fast, runs fast and works across episodes, so it generates trajectories efficiently for VLA.
Easy to develop through BFS.
Suitable for RSI: white box, efficient and controllable.
current code
carry
hang it
≈ 5 h3 rounds, branches at once≈ 26 h15 runs, one after another
@article{hu2026embodied,
title = {Embodied Turing Machines: Stateful Code
for Robot Recursive Self-Improvement},
author = {Hu, Kairui and Hu, Siyuan and Hong, Fangzhou
and Chen, Zhaoxi and Liu, Ziwei},
journal = {arXiv preprint arXiv:2610.12369},
year = {2026}
}