preprint · 2026-10-21
Toward Recursive Self-Improvement in Physical AI
On this page
Why it matters here
Luo and colleagues survey how experience can change robot policies, world models, external knowledge and learning procedures. They distinguish better task behavior from an improver learning to produce better later revisions, and propose controlled comparisons on new failures. Within the physical studies they review, the authors do not identify a controlled gain attributable to recursion. This is attributed methodological context, not a new robot experiment or a change to the tracker's autonomy assessment.
Recursive improvement evidenceEvaluation integrity
What to keep in mind
- A narrative preprint, rather than a new experiment or an independently exhaustive review. Cited studies and their reported results were not independently audited for this entry.
- The authors' conclusion concerns the physical studies they reviewed. It is provisional and does not establish that physical recursion is absent everywhere.
- The proposed controls require matched learners, held-out failures and protected outcome checks. Withholding learned knowledge is harder when it is inseparable from model weights.
- Human supervision, resets, hardware changes, safety and resource costs remain material. Autonomous morphology design and physical self-reconstruction lie outside the survey's scope.
- Checked scope: source metadata, introduction and scope, Sections 2.2 and 4.4, Figure 9 caption and conclusion. Full figure pixels and all cited experiments remain unchecked; no figure or numeric result is adopted.