X Square Robot Unveils WALL-SS World Model to Give Robots a Virtual Training Ground

The release follows the World Humanoid Robot Games in Beijing, where 2,056 robots from 666 teams across 16 countries competed in events ranging from track and field to real-world scenarios like household chores, hotel service, and emergency rescue. As robots move from arenas into homes, the stakes of trial-and-error rise—a spilled cup or a knocked-over chair is no longer a laughing matter. The challenge is that real environments lack fixed tracks and uniform props; a cup moved slightly, or a cloth left on a table, can invalidate a trained action.

WALL-SS tackles this by letting robots pre-play actions. It can simulate a continuous 60-second task such as pouring water or tidying objects, testing how different grasping approaches lead to missing, knocking over, or successfully lifting a cup. According to the company, strategies that perform well in WALL-SS also tend to perform well on physical robots, making the model useful for screening action policies before real-world deployment.

Unlike some world models that merely generate visually smooth videos without respecting physical causality, WALL-SS aligns action and visual outcome on the same timeline. It first builds a low-resolution preview and then refines it layer by layer, a process the authors call "next-scale autoregression." Given a fixed initial frame, changing the action trajectory changes the generated future, preventing the common "magnet-like grasp" error where objects appear to stick to colliders even with a visible gap.

In benchmark tests, WALL-SS scored 0.290 on action sensitivity, far above Cosmos3's 0.044 and other models' 0. It also achieved the highest trajectory accuracy of 0.539 among compared models, indicating that the simulated arm both follows the given motion and stays on path. The model incorporates failure data—such as missed grasps, slips, collisions, and recovery—to avoid conflating task goals with inevitable outcomes.

For long tasks, WALL-SS uses "scale-compressed long-term memory" to keep recent events detailed while compressing earlier history into a summary. This prevents extended prediction from drifting. The model also employs "per-scale dream forcing" by injecting disturbances during training, teaching the system to recover from small errors rather than letting them cascade into task failure.

To further bridge virtual and real gaps, WALL-SS uses two "referees"—one checks action consistency, another audits long-term coherence—to guide generation through an online policy alignment process. This improved action following from 0.264 to 0.290, trajectory precision from 0.512 to 0.539, and reduced cross-boundary error from 0.118 to 0.104.

The team tested whether virtual success predicts real success. Using five WALL-WM policies at different training stages, they ran 600 paired trials across six tasks with matched initial states. In 527 pairs, the final success matched between WALL-SS and physical robots. When comparing which of two strategies was better, the model agreed with real-world results 89% of the time. The correlation between virtual and real success rates across 30 task-policy combinations was 0.926, with a mean absolute error of 0.062.

The system also includes an action expert that can directly control a real robot. In a tabletop dual-arm task, WALL-SS achieved an average task progress score of 69.1, compared with 49.6 for π 0.5, 44.1 for DreamZero, and 34.0 for LingBot-VA.

While WALL-SS improves robotic cognition and training, the authors note that precise physical execution still depends on joint accuracy, dexterous hands, tactile sensing, and force control. The world model can weed out poor strategies early, but the last millimeters of contact remain a hardware and software co-design problem. As embodied intelligence moves from demos to duty, the ability to test and refine skills in a reliable virtual environment may become a decisive advantage.