Autonomous Robot has released its first next-scale autoregressive world model, WALL-SS, designed to address the challenge of action-result causal inconsistency in robotic world models. Through a hierarchical prediction mechanism that progresses from coarse to fine, WALL-SS ensures that distinct actions correspond accurately to distinct outcomes, enabling continuous reasoning for up to 60 seconds. Testing shows that WALL-SS achieves an action-following score of 0.29, significantly outperforming Cosmos3-Nano’s 0.044. After completing 600 simulated-to-real paired experiments, WALL-SS achieved a task success correlation coefficient of 0.926, demonstrating that the world model can help robots validate strategies in virtual environments prior to real-world deployment.

Article author and source: GeekPark

The next competition in robot world models may not be about visual quality, but about architecture.

Author | Li Yuan

Edited by Zheng Xuan

The gripper of a robotic arm isn’t actually holding the cup, yet the cup rises along with it as if attracted by a magnet. This seemingly absurd issue is troubling robotic world models, and the industry has even dubbed it “magnetic grasping.”

Video generation models only need to make the visuals appear plausible to succeed; but world models cannot simply aim to "look real." Most robot training data comes from successful demonstrations, causing models to develop shortcuts: rather than understanding subtle differences in actions, they directly generate the most likely outcome based on the visuals.

In the real world, a displacement of just a few millimeters in the gripper's position could change the outcome from successfully picking up a cup to knocking it over or missing it entirely. What the world model truly needs to understand is the causal relationship between actions and their outcomes.

On August 27, Independent Variable Robotics released WALL-SS, the next-scale autoregressive world model, aiming to transition world models from “functional” to “user-friendly.” Unlike generating an entire scene at once, WALL-SS depicts the future in a coarse-to-fine manner, ensuring that each action corresponds accurately to its outcome. It also integrates a long-term memory mechanism to enable continuous reasoning for up to 60 seconds.

Test results show that after 60 seconds of continuous inference, WALL-SS maintains visuals and motion trajectories closer to real video, achieving an action-following score of 0.29, compared to 0.044 for Cosmos3-Nano and 0 for all other tested models. WALL-SS is among the few models capable of generating predictions that follow action commands.

Zi-variable also announced the release of WALL-SS. For robotic world models that have not yet established a unified technical roadmap, the significance of this release goes beyond opening a single model—it invites the industry to collectively verify whether world models are making predictions based on actions, and whether virtual simulations can truly serve real-world robots.

When the technological paradigm has not yet converged, continuous exploration itself is a competitive advantage. Embodied intelligence requires not only hardware that can be rapidly deployed, but also companies that persistently challenge existing approaches and push the boundaries of core technologies.

World models are the main track toward embodied intelligence.

In 2026, world models are becoming one of the most prominent technological directions in robotics.

NVIDIA continues to iterate on Cosmos and is beginning to apply video world models to robot control and planning; Meta’s V-JEPA 2 attempts to enable robots to perform grasping in unfamiliar environments by predicting changes in the physical world. An increasing number of studies are also exploring the use of world models to replace part of real-world testing, allowing robots to predict the outcomes of different strategies in a virtual environment before taking action.

Understanding this trend isn't complicated. Large language models learn language patterns by predicting the next word, while world models predict the next moment of the world: if a robot moves left, will the cup tip over; if the gripper tightens slightly more, can it grasp the object; if an action is repeated for ten seconds, what will the scene ultimately look like?

This is especially important for embodied intelligence. Over the past few years, the industry has primarily focused on solving the question of how robots "do" things, with VLAs directly outputting actions based on visual and language instructions. But as robotic capabilities improve, a new issue is becoming increasingly prominent: how to determine whether a given strategy is truly effective.

If each model iteration relies on real-world testing, the cost is high and the cycle is long, easily becoming a bottleneck for training and iteration. Traditional simulation can handle some testing, but it requires manual modeling of object materials, friction, deformation, and contact relationships, making it difficult to capture the complexity of the real world. World models, on the other hand, aim to learn how the world changes from real video and robot interaction data, enabling robots to repeatedly predict and validate strategies in a virtual environment.

It can even further participate in the robot’s decision-making. After seeing the cup, the VLA directly decides to “reach out and grab,” while the world model can first predict the consequences of different actions, helping the robot choose a more suitable strategy—transitioning from “acting directly” to “imagining first, then acting.”

However, to truly become the infrastructure for embodied intelligence, today's world models still face three key barriers.

First is the causal consistency between action and outcome. The model must not only generate visually expected scenes but also accurately reflect the results caused by the action; otherwise, a strategy that cannot actually grasp a cup might still be deemed successful in the virtual world.

Second is long-horizon reasoning capability. Robot tasks often last tens of seconds or longer, requiring the model to continuously use its own predictions as inputs for the next step, where even minor errors can accumulate over prolonged reasoning.

Third is the consistency between virtual simulation and real-world operation. No matter how realistic the generation or how long the extrapolation, it does not mean the model can truly evaluate robotic strategies. Only when performance in the virtual world can be stably correlated with real-world outcomes does the world model truly gain value as a substitute for partial real-world trial and error.

These three barriers are also becoming the core of the next phase of world model competition.

WALL-SS: Enhancing the world model from the architectural foundation

WALL-SS is a new attempt to address these issues. According to the paper, it is the first world model to employ a next-scale autoregressive architecture.

Previously, many world models followed the video Diffusion approach: inputting historical frames and actions, then generating future videos through multiple denoising steps. Although actions served as conditioning inputs, they were not explicitly encoded into the "action → outcome" generation chain, causing models to rely heavily on visual priors and directly predict a "likely successful" future—even if the gripper did not actually grasp the cup, the model might still generate a frame showing the cup being successfully lifted.

WALL-SS organizes observations and actions into causal sequences of "observation—action—new observation," enabling actions to genuinely influence the generation of future states. In simple terms, it first determines how the world changes at a high level, then progressively refines the details: coarse-grained levels establish the approximate states of the robotic arm and objects, while fine-grained levels further reconstruct motion and contact. Combined with failure, takeover, and correction data, the model learns that "different actions lead to different outcomes."

This capability is validated by the metrics: WALL-SS achieves an action follow score of 0.29 and a trajectory accuracy of 0.539; in comparison, Cosmos3-Nano achieves only an action follow score of 0.044. Other models score exactly 0.

Another key issue is whether the world model can “remember long enough.” WALL-SS preserves recent details through multi-scale long-term memory, compresses distant historical data, and enables the model to continue simulating from its own predictions during training, reducing the accumulation of errors. This allows the model to perform continuous simulations for up to 60 seconds, whereas other models are typically limited to 20–30 seconds.

More importantly, these capabilities were ultimately validated on real hardware. The autonomous system executed the same strategies on both WALL-SS and real robots, completing 600 simulated-to-real paired experiments. The results showed a correlation coefficient of 0.926 in task success rates between the two settings, and an accuracy of 89% in ranking the relative strengths of different strategies.

This means world models are beginning to offer a more significant value: not only generating futures that appear realistic, but also helping robots determine the outcomes of different actions and validate strategies in advance within a virtual environment.

Of course, WALL-SS currently has limitations. Its action inputs primarily consist of the end-effector position, orientation, and gripper opening/closing, excluding full joint states, forces, and torques; therefore, support for fine-grained contact and multi-finger dexterous hands still requires further validation.

Continue innovating in architecture until the roadmap converges.

WALL-SS is not a sudden result of chasing the trend of world models. Over the past year, Independent Variable has been exploring the same question: how to truly connect a robot’s understanding of the world with its predictions of actions and their actual execution.

WALL-OSS integrates vision, language, and robotic actions into a single embodied foundation model through large-scale multimodal pretraining; WALL-OSS-0.5 further advances the transfer of pretraining capabilities to real robots; subsequently, WALL-WM begins jointly modeling world dynamics and robotic actions, while X-Tokenizer explores how diverse robots and actions can be unified within a single model. With WALL-SS, this path further points to a core question: What exactly happens to the world after executing an action?

Running throughout this series of explorations is a long-term commitment to modeling the unified "world-action" space. Robots must not only be able to act but also anticipate the consequences of their actions and continuously refine their behavior based on outcomes. As models, data, ontologies, and real-world experience accumulate, these explorations gradually form a self-reinforcing technical feedback loop.

The foundational architecture of embodied intelligence is far from converged. At this stage, what truly matters is not whether each attempt correctly predicts the final answer, but whether new architectural hypotheses can be continuously proposed, validated with data and real-world systems, and iteratively refined based on the results.

This announcement of WALL-SS continues this line of thinking. What is being released is not just a model, but a set of technical approaches that the industry can jointly validate: Is the next scale of autoregression better suited for robotics? Can actions and outcomes be truly aligned? Can virtual testing accurately reflect real-world performance? Researchers can build on this to further explore these questions, while robotics companies can directly test whether it fits into their own training and evaluation systems.

This also highlights the greater significance of this release for the early-stage embodied intelligence industry: reducing the cost of industry exploration and enabling different approaches to be tested in real-world applications.

Over the past few years, competition in domestic embodied intelligence has primarily focused on the physical body, data, and product implementation. However, an industry that has not yet stabilized cannot simply wait for a mature architecture to emerge before catching up—it also requires pioneers to engage at more fundamental levels, such as world models, action representations, and learning paradigms, to continuously explore the next possibility.

WALL-SS is certainly not the final answer, but it represents a shift: domestic companies are beginning not just to follow existing paths, but to propose their own architectural assumptions and open them up to industry scrutiny.

When a technological shift truly occurs, what defines a company is not whether it kept up with the previous generation, but whether it has consistently maintained the ability to explore the next generation.