Can a robot improve itself? A six-level framework concludes that physical recursive self-improvement does not yet exist

Researchers propose a scale from L0 to L5 to characterise the self-improvement of robots and note that no published work actually reaches the recursive stage.

Robots perform their tasks more and more autonomously, but their progress is still largely dependent on the humans who retrain them. A study published on arXiv is interested in what its authors call the "physical RSI": a recursive self-improvement where physical experience permanently alters both the robot's capabilities and its very way of learning, so that each cycle of improvement becomes more effective than the previous one.

To be clear, the team proposes a scale of L0 to L5 that distinguishes the capacity updates from those on the learning method, and the earnings limited to one task from those that transfer to others. The L5 level is reserved for an independently verified recursive gain. The physical world complicates exercise: noise observations, continuous dynamics, costly trials and sometimes irreversible consequences.

After reviewing existing methods of adapting the execution environment to model training and safety mechanisms, the authors conclude that none of them yet demonstrate this recursive self-improvement within the meaning of their definition. They set four priorities: reliable returns, causal attribution of progress, preservation of safety through common updates and test benches.