Toward Recursive Self-Improvement in Physical AI
Meng Luo, Zechen Bai, Shanqing Xu, Yanlin Li, Kaixin Li, Mengchen Qian, Weifeng Lu, Ran Zhang, Yingtian Zou, Ziyang Luo, Wei Chen, Hongzhan Lin
cs.RO, cs.AI
2026-10-21
A five-institution survey grades robot self-improvement claim by claim: behavior, later learning, and procedure improvement have hardware evidence; the recursive links have none.
Recursive self-improvement (RSI) means a system improves the process by which it improves, not only its behavior. Digital agents already gesture at this: STOP turns a code-improving program on its own source, and the Darwin Gödel Machine lets improved coding agents modify themselves, with executable tests judging every candidate. Physics changes the deal in four ways. Outcome checks rely on sensors, people, or learned detectors that can be wrong, and a scoring error can be reinforced and then executed on hardware. Every candidate costs robot time, resets, supervision, and wear. Practice changes its own conditions, since grippers wear, calibration drifts, and objects move, while rolling back software does not repair a damaged object. And the things open to improvement now include controllers, dynamics models, and the interface to a particular body.
The more immediate problem is term inflation. Plenty of robot systems now call themselves self-improving, self-evolving, or even RSI, while several rungs between better behavior and recursion have never been tested. This NUS-led survey from five institutions, with a companion GitHub reading list, starts by fixing the language: each claim is defined by the comparison that would establish it.
The framework sorts every retained change by where it lands: the policy (what executes), the world model (prediction and practice), external knowledge (reuse and recovery), or the learning procedure (what experience to collect, how to judge it, how to update). The first three change what the robot does; the fourth changes how it learns, and recursion can only start there.
On top sits a ladder of four claims, each tied to its own comparison. Behavioral improvement requires the new version to win a common task test. Later-learning benefit requires retained data, models, or skills to make a subsequent learner better. Procedure improvement requires a revised collector, evaluator, or updater to teach better on matched learners. Recursion adds two links: R1, repair knowledge from earlier fixes improves the improver itself; R2, the improved improver helps produce its own next improvement. Two controls carry the verdict: a frozen-knowledge control withholds only what the improver acquired, and a fixed-improver control removes the improvement step entirely. A fixed improver can keep improving procedures forever while satisfying neither link.
The survey also names four conditions for a gain to carry forward: the feedback addresses the remaining learning need, the receiving component can absorb it, its value survives changes in task or hardware, and continued practice stays safe and affordable.
The verdict: the first three claims have physical evidence, concentrated in manipulation; the two recursive links have zero controlled demonstrations. Data volume and component accuracy are unreliable proxies for progress. Representative numbers, all from the studies reviewed:
| System / study | Setting | Result |
| ConRFT | 8 real tasks, online fine-tuning | 96.3% average success after 45–90 min |
| VLAC | 4 real tasks | 30% to 90% within 200 episodes |
| Max-Q selective imitation | real USB insertion | 99% in half an hour; HIL-SERL needs 5 h to its 100% ceiling |
| PLD vs RLPD | matched collection, same distillation | distilled generalist 30/30 vs 16/30 trials |
| PhysRSI-eval ablation | GPT-6 Astra, rounds 4–6 | withholding history drops 15/36 to 3/36 |
| FailBench | 2,197 real and simulated attempts | best of 13 VLM detectors: 0.77 balanced accuracy |
Two findings cut against the more-data intuition. Mirchandani et al. find that added autonomous data yield modest gains that saturate and can regress over two to four rounds. DexTouch-WM's world model gets better as a surrogate evaluator, yet replacing half the real training trajectories with its synthetic ones lowers physical scores for all three policy families tested. How often a collector succeeds says nothing about the training value of what it collects.
For anyone building self-improving robots, the value here is experimental discipline: which two controls to run, which variables to match, which failures to hold out before claiming recursion, written as an executable checklist, down to a two-round experiment design with a hardware cost estimate. For benchmark builders, Section 2.5 maps the landscape from LIBERO to PhysRSI-eval and names the gap: most suites score a system at one point in time, and cross-round protocols barely exist. This is a position and survey paper with no new system. What it does is drag a concept bloated by marketing talk back onto falsifiable ground.
The authors state them plainly: the reviewed studies concentrate in manipulation, with navigation, locomotion, and flight supporting only isolated mechanisms; recursive gains still need dedicated verification; the four conditions are diagnostic lenses, not validated constructs. The R2 control has no clean design when the acquired knowledge lives in the improver's weights, and the paper only sketches one. Two things to press on: the four-claim ladder is this paper's own operationalization, and a concurrent survey (Zhou et al., L0–L5) grades levels differently, so what counts as recursive will stay contested. Most cited numbers are within-study comparisons, with tasks, budgets, and metrics too uneven to compare across studies, as the survey itself warns. It cannot answer whether recursion is achievable on hardware; it puts that question in testable form for the first time.