Reproduce “Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kairui Hu, Siyuan Hu, Fangzhou Hong, Zhaoxi Chen, Ziwei Liu
COAP represents scene state in code and uses a fixed task program to choose actions, check outcomes, and update its records. At test time, it uses camera images and robot measurements without a model service. On RoboDojo, it reaches 70.24% overall success, versus 31.38% for PhysicalRSI. Results are from simulation and depend on state measurement and code robustness.
Key insights
- Explicit state and persistent episode memory let a fixed program track objects and subgoals, including when objects become occluded, and make policy decisions easier to inspect.
- Reusable code supports task composition: the paper reports that 83% of the code executed by a new task comes from the shared library.
- The approach depends on reliable state measurement; the authors attribute 62% of failures to measurement problems and 37% to incomplete logic or code design.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.