arxiv:2610.29677

Toward Recursive Self-Improvement in Physical AI

Published on Oct 21

· Submitted by

Hongzhan Lin on Oct 23

National University of Singapore

Upvote

Authors:

Meng Luo ,

Abstract

Recursive self-improvement (RSI) describes a system that improves not only its behavior but also the process by which it improves. Recent digital agents test this idea by rewriting their own code and checking each version with executable tests. Bringing these mechanisms into the physical world requires learning from partial observations through costly trials, while accounting for changing operating conditions and potentially irreversible consequences. We position RSI within robot learning and trace pathways from improving robot capabilities to improving the processes that generate them. We examine how physical experience can strengthen these improvement mechanisms and when the resulting gains can support further self-improvement of the mechanisms themselves. Gains carry forward only when feedback addresses remaining learning needs, the receiving component can use the change, its value survives changes in tasks or hardware, and continued practice remains safe and affordable. Existing studies, concentrated in manipulation, support behavioral improvement, benefits to later learning, and procedure improvement. Some systems also change the improver's knowledge or tools, but the physical studies reviewed here have yet to isolate gains attributable to recursion. We synthesize these findings into evaluation principles and research directions for machines that, trial by trial, safely improve the way they improve.

View arXiv page View PDF Project page GitHub 15 Add to collection

Community

danielhzlin

Paper submitter about 23 hours ago

Toward Recursive Self-Improvement in Physical AI

Recursive self-improvement (RSI) asks for more than a better robot: the process that produces improvements should itself get better. Digital agents already rewrite their own code and check each version against executable tests. But what does it take to carry these mechanisms into the physical world, where feedback is partial, trials are costly, and some mistakes cannot be undone?

What the paper covers:

🧭 A ladder of claims: behavioral improvement → later-learning benefit → procedure improvement → recursion. Each step is defined by the comparison that would establish it.
🤖 Where improvement lands: policies, world models, external knowledge (skills, code, memory), the learning procedure, and the improver that revises it.
🔁 When gains carry forward: four conditions. Feedback addresses the remaining need, the receiving component can use the change, the change stays useful under task or hardware shift, and practice remains safe and affordable.
📍 Where the field stands: existing robot studies show behavioral improvement, later-learning benefits, and procedure improvement, but none yet isolates a gain attributable to recursion.
🧪 What's still open: fair tests the system can't overfit to, gains that keep stacking up, safe practice while the procedure changes, and what a first real demo of physical RSI would need to show.

🌐 Project page: https://physical2rsi.github.io/
💻 Paper list: https://github.com/Physical2RSI/Awesome-Physical2RSI

Feedback and pointers to work we missed are very welcome!