Shengshu's new world model explores RSI to let robots self-evolve

Shengshu has introduced a new world model and pointed it at RSI, the idea that a system keeps improving itself without a human intervening at every step — in short, letting robots begin to evolve on their own, according to Chinese tech outlet QbitAI.

A world model acts as a rehearsal space for robotics. It learns how the environment changes in response to actions, then generates large amounts of virtual experience inside the model, which is cheaper than trial and error on real hardware and can cover edge cases that are hard to stage in the physical world.

RSI pushes that logic one step further. If the world model answers where training data comes from, RSI addresses who sets the next training goal: the model generates its own tasks, evaluates its own results and locks in whatever proved useful, closing the loop.

That is also where the approach invites skepticism. A self-improvement loop without dependable external checks tends to reinforce its own preferences, and in robotics the gap between simulation and reality makes that risk sharper — metrics that climb steadily in a virtual environment do not always transfer to a physical machine.

So the next questions are about verifiability: whether Shengshu publishes how the model is evaluated and how tasks are designed, whether self-evolution reproduces on real hardware, and whether the method settles into general capability or only works on a narrow set of tasks.