If a robot could “rehearse” the future before taking action—anticipating what contacts will occur, how objects will move, and whether its trajectories are feasible—would it become smarter? 
 

This is precisely the question that RoboScience’s “Embodied World Model” seeks to address. It does not merely generate aesthetically pleasing videos; rather, it enables robots to construct an interactive, predictable, and physically consistent “imaginative space” in the digital world, allowing for endless trial and error, rehearsal of the future, and risk assessment—culminating in confident execution in the real world. 

Currently, most world models in the field focus on two main directions: one is 2D video generation, which predicts the next frame’s pixels but lacks an understanding of three-dimensional space; the other is 3D static reconstruction, which can recover spatial structure but cannot predict how objects will move over time. RoboScience’s “Embodied World Model” takes the third approach—3D dynamic world modeling: predicting the continuous trajectories of objects over time in three-dimensional space.

The “Embodied World Model” is one of the two core engines of RoboScience’s VLOA large-scale model, and together with the “General Manipulation Model,” it forms a complete closed loop on the path to general-purpose embodied intelligence. The former is responsible for understanding the physical world and simulating future trajectories, while the latter translates imagination into precise actions. 
 

Today, we’ll begin by lifting the veil on the first half.

Our “Embodied World Model” takes natural language instructions and visual images as inputs, can process both single-view and multi-view imagery, and accurately localizes target objects and predicts their future motion states in complex scenes. 
 

Unlike traditional methods that predict the next frame’s pixels, the “embodied world model” focuses on the semantic changes in object states— The object’s position, orientation, deformation, and its interactions with the surrounding environment in three-dimensional space. It decomposes the scene into individual objects and predicts their future motion trajectories separately.

The final output is a 3D point cloud trajectory of the object’s future motion path:  A sequence of 3D points with timestamps, where each point includes position coordinates, pose information, a time step, and a prediction confidence score.

Why 3D point clouds? Because it is explicitly interpretable – it allows for an intuitive visualization of the model’s predicted trajectory; by modeling in real 3D space, it naturally satisfies geometric constraints; moreover, the trajectory can be directly used as input to downstream manipulation models, enabling seamless end-to-end transfer from perception to execution.

Videos generated by the “Embodied World Model” and their 3D point-cloud trajectories 
 

These 3D point-cloud trajectories are not generated out of thin air; rather, they are end-to-end produced from input visual images and language instructions using a neural network architecture specifically designed for dynamic three-dimensional environments. 
 

The figure below illustrates the internal structure of the “Embodied World Model”: 
 

The “Embodied World Model” first encodes RGB observations, 3D point‑cloud priors, and task instructions into computable semantic and spatial representations, then feeds these into a world‑causal Transformer to model the future evolution of the world under the given task conditions, thereby constructing a unified latent world representation. Subsequently, the model generates 3D flows for the scene and target objects via a decoding process, and can further produce future manipulation videos along optional branches, thus linking “seeing the present, understanding the instruction, and predicting the future” into a seamless end‑to‑end pipeline. 
 

This architecture endows the model with three key highlights:

▎Multi-object grasping demonstration 
 

In the same storage scenario, the model generates tailored grasping and presentation motions for various objects, such as cotton swab boxes and bottles.

Facing the lemon tea box, coffee capsule container, orange soda bottle, and bagged coffee on the tabletop, the model generates precise motion trajectories for each object. 
 

 

Highlight No. 2:

Dynamic Process Modeling — Imagining Physical Changes

 

Given the first frame of a first-person perspective, the model can “imagine” the entire subsequent pouring process—how the kettle tilts, how the water flows into the cup, and how the water level in the cup rises.

 

Even when the act of pouring water involves fluid dynamics and fine‑grained manipulation, the model can still generate 3D point‑cloud trajectories that adhere to physical laws. This capability goes far beyond mere “video completion”; it represents a genuine modeling of future physical phenomena.

 

▎First-Person Perspective Operation Demonstration 
 

The model predicts the entire process of a kettle pouring water into a bowl, including the tilt angle, the flow of water, and the rise in water level.

The model predicts the motion trajectory of a handheld white mug as it is placed on a dinner plate.

The model not only recognizes objects but also captures semantic distinctions in instructions—identifying who the subject is, what the action entails, and how the intended meaning varies. This reflects its capability for cross-modal semantic alignment and fine-grained instance-level differentiation.

The model generates distinct actions for the robotic arm: placing a white mug and a small green bowl containing food into separate compartments of an orange bowl.

The model predicts the robotic arm’s actions of retrieving brown garments and fluorescent yellow garments from the laundry basket and placing them into the washing machine.

Through the visualization examples above, we have transformed the world model from a “black box” into an interpretable, debuggable, and trustworthy cognitive engine. The trajectory changes in each video directly reflect the model’s internal reasoning.

The capabilities demonstrated in the aforementioned cases—such as cross‑object generalization, dynamic process modeling, and instruction following—are underpinned by four core technical attributes inherent to the model. These attributes ensure that the “embodied world model” is not merely “imagination,” but rather “reliable imagination.” 
 

· Physics constraint satisfaction: All trajectories strictly adhere to real-world physical constraints, such as dynamics, collisions, and stability. In the water-pouring example, the alignment between the kettle’s tilt angle and the water‑flow trajectory, along with the smooth rise of the water surface, demonstrates the model’s precise grasp of gravity and fluid behavior. This is something 2D video generation cannot achieve—because in a 2D world there is no defined direction of gravity, whereas our model genuinely “understands” the laws of physics in three-dimensional space.

· Native support for physics-based multi-solution modeling The real world is inherently uncertain. This approach leverages the generative capabilities of diffusion models to construct probabilistic distributions of physical evolution in the latent space, thereby enabling the derivation of multiple plausible trajectory plans for the same task. This capacity to model uncertainty provides a robust foundational support for ensuring decision‑making safety in complex environments within embodied intelligence.

· Long-term temporal spatial consistency: In complex, multi-step tasks, the model maintains global temporal and spatial consistency in its predictions. In long‑duration water‑pouring video prediction, the relative positions of objects remain physically plausible throughout, with no hallucinations. 
 

· Hardware decoupling: The model’s core is decoupled from the specific robot architecture, enabling the generated plans to be seamlessly transferred to any robotic platform—whether a robotic arm, a humanoid robot, or a dexterous hand—while all of them can interpret the same object trajectories. 
 

It is precisely these four core technological features that ensure every “imagination” generated by the Embodied World Model is grounded in evidence and governed by clear principles. 
 

The capabilities of the “embodied world model” also continue to evolve with each iteration of training. The figure below illustrates how, during pre-training, the model’s performance improves across several key metrics as the number of training iterations increases.

Figure 1. Trends in metric values during model training iterations

Content Alignment, Subjective Quality,

Photometric Consistency and Motion Smoothness

They all continue to improve throughout the fine-tuning process. The ⭑ symbol denotes the final checkpoint,

The annotation provides the final score and the relative improvement over the initial model. 

Figure 2. Performance gains from the base model to the final checkpoint

The left figure shows the absolute increase for each indicator, while the right figure presents the relative percentage increase.

The improvement in photometric consistency is the greatest,

Next is Motion Smoothness. 

As you can see, the more data we feed into the model and the greater the computational resources we allocate, the more accurately the model understands the physical world, and the closer its predicted future trajectories are to reality.

This is precisely the Scaling Law in the field of embodied intelligence: model growth is both predictable and sustainable. As we continue to expand our video dataset at a pace of hundreds of thousands of hours per week, the capabilities of our world model will keep advancing, providing robots with an increasingly reliable “mental simulation” capability.

In the full VLOA architecture, the embodied world model serves as the “cognitive brain”—understanding the physical world, predicting object states, and generating executable 3D point-cloud trajectories. This trajectory is then passed through object trajectory Interface, passed to the next core module: General Manipulation Model. 
 

It is worth noting that the underlying foundation underpinning the continuous evolution of these two major models is a data system that places equal emphasis on scale and quality.

 

Through a fully automated data annotation and cleaning pipeline, we extract high‑value content from vast volumes of internet‑hosted video—focusing on object state changes and physical interactions. To date, we have amassed over one million hours of high‑dimensional, multimodal operation‑related video data (comprising tens of millions of video clips), with growth continuing at a pace of hundreds of thousands of hours per week. Our goal is to build a world‑leading video dataset exceeding ten million hours by the end of 2026, providing an inexhaustible source of training data to fuel the ongoing evolution of our embodied world model.

 

Meanwhile, in terms of “General Manipulation Model” data, leveraging our proprietary multimodal physics engine, we have already amassed a high-quality, full‑space object manipulation dataset comprising 10 billion instances, with the goal of building a manipulation dataset exceeding 1 trillion instances by 2026.

 

Today, we demonstrated how an embodied world model can open the black box of physical cognition by leveraging 3D point‑cloud trajectories. Yet this is only half the story—how do we translate these imagined trajectories into precise contact points on a robotic hand, appropriate force control, and smooth, fluid motions? And how do we adapt to robots of diverse forms and configurations?

 

This is precisely the answer that will soon be revealed in “VLOA Series Explained (Part 2): The General Operational Model.”

 

And underpinning these capabilities are the robot hardware products we’ve developed in parallel. They serve as the ideal platform for VLOA’s large‑model technology and represent the ultimate form of intelligent systems as they are deployed in the physical world.

 

Stay tuned.