X-Square's WALL-SS World Model Enables Reliable Transfer from Virtual to Physical Robot Tasks

APPSO · wechat · 2026-08-27

World models aim to reduce robot trial-and-error costs via virtual simulation, but often struggle with issues like "magnetic grasping" and long-term memory drift. X-Square Robot's autoregressive world model, WALL-SS, addresses these by using "next-scale autoregression" to ensure actions actually alter outcomes, "scale-compressed memory" for long-horizon consistency, and "dream forcing" with "visual dynamics online policy alignment" to improve strategy selection reliability.

In evaluations, WALL-SS significantly outperformed baselines in action following and trajectory precision. Experiments showed a 0.926 correlation between virtual strategy rankings and real-world results, proving that virtual performance can effectively guide real-world testing. Combined with an action expert, the system excels in dual-arm desktop tasks, offering a more efficient path for embodied intelligence to move from demos to real-world applications.

Original post →

More from Embodied

Embodied channel →