If a robot can “rehearse” the future, how should it turn its imagination into reality? This is precisely the central question that the “General Manipulation Model” within RoboScience's VLOA large-scale model seeks to address.

In the previous installment, VLOA’s “embodied world model” used 3D point‑cloud trajectories to demystify physical cognition—enabling robots to anticipate how objects will move, make contact, and deform. Yet no matter how precise the simulation, without a reliable pair of “hands” to carry it out, it all remains little more than castles in the air.

As the second-largest core engine of VLOA, the “General Manipulation Model” has precisely the following mission: It receives 3D point-cloud trajectories generated by the world model, converts them into contact points, force‑control commands, and joint‑angle instructions that can drive any robot, ensuring that every “simulation” is precisely replicated in the physical world.

Currently, manipulation models commonly face three major bottlenecks: a generalization challenge—losing effectiveness when presented with a different object; an inability to handle fine-grained manipulations; and the accumulation of errors in long-horizon tasks, where a single mistake can cascade into subsequent steps. Most manipulation models in the field are built on “atomic skill libraries,” which decompose tasks into discrete skills such as grasping and placing, each supported by a dedicated model. This fragmented approach suffers from poor scalability and struggles to adapt to new tasks.

RoboScience’s “General Manipulation Model” is a large-scale model with over one billion (1B) parameters, jointly trained across all skills to produce a unified manipulation representation. It eliminates the need to train separate sub‑models for new objects or new actions, leveraging shared physical common sense and trajectory priors. This enables general-purpose manipulation capabilities across objects, tasks, and robot platforms. 。 This model has been redesigned from the ground up, leveraging an efficient closed-loop workflow—“physics engine–simulation data–end-to-end training”—to systematically address the challenges of generalization and dexterous manipulation, enabling robots to achieve truly versatile manipulation capabilities.

The input to the “General Manipulation Model” is precisely the output of the “Embodied World Model.” Object Trajectory , namely a sequence of 3D object point clouds and environmental point clouds, each timestamped, which describe the object’s future position, pose, and deformation, as well as changes in the grasping environment.

The general-purpose manipulation model achieves an inference speed of over 3 fps and enables closed-loop control of robot joint angles based on point-cloud inputs from objects and the environment. Unlike traditional approaches that rely on large amounts of paired motion data, our model is trajectory-conditioned: it does not need to relearn “where to go,” but rather learns “how to get there” and how to manipulate objects.

The trajectories generated by the world model already incorporate rich geometric and physical priors, and the manipulation model simply needs to translate them into low-level control signals. This endows the model with remarkable generalization across objects and scenes.

“General Operations Model” Architecture Diagram 
 

This architecture endows the model with three key highlights:

When confronted with objects of varying geometries and physical properties, the model can instantly recognize their three-dimensional shapes and associated physical parameters, automatically selecting optimal contact points and gripping forces to generate tailored grasping strategies. Whether the objects are individually placed on a tabletop or nestled in a storage bin, bowl, or dish—amid cluttered, densely packed arrangements—the model consistently achieves stable grasping.

▎ Dexterous grasping in arbitrary cluttered environments

The figure below illustrates the model’s differentiated grasping performance across two types of cluttered scenes:

Desktop objects: When encountering an irregularly shaped fawn toy—with protruding antlers—the robot automatically detects its geometric features and selects a stable grasping point on the body, avoiding damage to the decorative elements, then executes the grasp with appropriately calibrated gripping force. For a bottled ketchup container, the robot grips the midsection of the bottle to maintain balance and prevent tipping or slipping.

After grasping, the model places objects into designated areas according to their categories, enabling a fully automated workflow from grasping to sorting. This capability demonstrates the model’s adaptability to complex scenes and crowded environments, as well as its deep understanding of the physical properties of diverse objects.

Inbox objects: When the square snack box is surrounded by other objects, the model adjusts its grasp to pick it up in a confined space without disturbing the surrounding items; when faced with an oddly shaped bottle opener resting in a bowl, the model autonomously reconfigures its dexterous hand to gently grasp the long handle and retrieve it from the bowl.

▎ Cross-embodiedment dexterous grasping

The model is fully decoupled from the robot hardware, allowing the same set of manipulation policies to be directly transferred to dexterous hands of different configurations—whether two‑finger, three‑finger, or five‑finger—without any modifications. The model first generates a 3D point cloud of the object via visual perception, accurately capturing its three‑dimensional geometry and spatial pose, and then adaptively computes the optimal grasping configuration and force‑control strategy.

Take the X‑hand and the LEAP Hand as examples: the two differ significantly in their mechanical design. The X‑hand employs a fully geared, direct‑drive transmission, with 12 active degrees of freedom, delivering a maximum grip force of 80 N per hand and capable of lifting a 25‑kg load. In contrast, the LEAP Hand utilizes direct‑driven actuation combined with a four‑bar linkage kinematic architecture, achieving a total of 16 to 20 degrees of freedom.

The two dexterous hands differ in their degrees of freedom allocation, actuation schemes, and dimensional specifications; however, a universal manipulation model can adaptively generate both a conformal grasping strategy for a green pepper and a precision grasping strategy for a watermelon slice on either hand.

Opening an envelope requires millinewton‑level insertion force, balancing a coin on its edge demands precise dynamic control, and picking up a potato chip must avoid crushing it; injecting a liquid with a syringe necessitates exacting control over both the injection speed and dosage. The model can stably perform all these tasks, which place extremely high demands on force‑control accuracy, contact‑force sensing, and real‑time adaptation.

The model integrates multimodal sensory signals, including vision, haptics, and force perception, and dynamically adapts in real time during execution:

▎ Detailed Operation Demonstration 
 

Standing a coin: A coin is balanced upright on the tabletop through dynamic equilibrium.   Opening the envelope: Precisely control the cutting angle and pressure to score along the edge without tearing the paper.

Grab the chips: Gently pinch the edges to avoid crushing and keep the chips intact.

Syringe injection: Precisely controls the infusion rate and force, enabling quantitative and stable fluid delivery.

Grasp seaweed/eggshell/ice cream cone: Gently pinch the edges to avoid crushing, keeping the nori sheet, eggshell, or ice cream cone intact.

The model possesses core capabilities for handling complex long-range tasks and dynamic closed-loop operations. The following three representative demos highlight its breakthroughs in multi-step planning, fine-grained force control, and environmental adaptability:

▎ Long-Range Task and Closed-Loop Operation Demonstration 
 

Furniture assembly: The model reads the instruction manual, autonomously decomposes multi-step tasks, and achieves high-precision plug‑in and rotational coordination through dual‑arm collaboration. By leveraging real-time force feedback to dynamically adjust its strategy, it can automatically recover and resume execution even if interrupted mid‑process.

Dynamic grasping on a conveyor belt: The model detects the velocity and pose of moving objects in real time, iteratively refines the grasping point and approach trajectory, and achieves stable grasping even as the object continues to move, adapting to velocity fluctuations and changes in orientation.

At the heart of this capability is the model’s ability to not only plan coherent multi-step actions but also to perceive, make decisions, and adapt in real time within dynamic environments. By employing a unified framework, it can analyze physical interactions across diverse scenarios—such as force feedback, deformation prediction, and motion planning—without requiring separate algorithms for each object or robot, thereby enabling true closed-loop “hand–eye coordination.”

Through the visualization examples above, we have transformed the general-purpose operation model from a “black box” into an interpretable, debuggable, and trustworthy execution engine. The action details in each video are direct reflections of the model’s real-time internal decision-making.

The capabilities demonstrated in the aforementioned cases—such as cross‑object grasping, fine force control, and long‑range execution—are underpinned by four core technical attributes inherent to the model. These attributes ensure that the “General Manipulation Model” is not only “flexible” but also “reliable.”

​
 

· Full-space object support: The model supports a wide range of manipulation tasks involving objects across the full spatial domain, including rigid bodies, hinged bodies, and 1D/2D/3D deformable objects, spanning the entire spectrum from rigid grasping to flexible deformation.

· Cross-ontology and Closed-loop Operations: The model is fully decoupled from the hardware, allowing the same set of policies to be seamlessly migrated across diverse platforms, including robotic arms, humanoid robots, and dexterous hands. It also supports closed-loop operation, continuously acquiring multimodal sensory data—such as visual, tactile, and force information—during execution and dynamically adjusting actions to maintain precision even in dynamic environments.

· Physics-based simulation closed loop: Through an efficient closed-loop comprising a physics engine, simulation data, and end-to-end training, the model undergoes large-scale pre-training in a virtual environment, acquiring a wealth of physical interaction skills. With just a small amount of real-world data for fine-tuning, it can be rapidly transferred to real-world scenarios, significantly reducing data-collection costs.

· Scaling Drives Evolution: Leveraging 10 billion high-quality manipulation data points accumulated through our in-house multimodal physics engine, the model’s success rate in tasks such as object grasping, fine manipulation, and long-range operations exhibits a predictable power-law improvement as the dataset size grows. By 2026, our goal is to build a 1‑trillion‑operation dataset, providing an inexhaustible source of training data to fuel continuous advancement.

Experiments demonstrate that as the model size increases, both the success rate and the grasp diversity—measured by the variance of joint angles in successful grasps—also improve.

At this point, the two core engines of the VLOA large model have been fully unveiled:

“Embodied World Model”: A 3D dynamic world model that uses the 3D point-cloud trajectories of objects to simulate future physical states. Behind it lies over one million hours of object-centric, high-dimensional multimodal video data—comprising tens of millions of clips—and this volume is growing at a rate of hundreds of thousands of hours per week, with the goal of building a globally leading video dataset totaling tens of millions of hours by 2026.

“General Operation Model”: Through a closed-loop system of “physics engine–simulation data–end-to-end training,” trajectories are translated into contact points, force control signals, and joint commands, enabling the grasping of arbitrary objects, fine manipulation, and long-range tasks. Underpinning this approach is 10 billion (10B) instances of full‑space object‑manipulation data accumulated via our in-house physics engine, with a target of surpassing 1 trillion (1T) by 2026.

The two systems seamlessly collaborate via the Object Trajectory interface, forming a complete intelligent closed loop that spans from perception to execution. From acquiring physical common sense through vast amounts of video and simulation data, to precisely simulating 3D point‑cloud trajectories, to reliably executing arbitrary objects, fine‑grained manipulations, and long‑range tasks in the real world—VLOA, a large‑scale model, is redefining the frontiers of general‑purpose embodied intelligence.

RoboScience is dedicated to building world‑leading embodied‑intelligence large models and foundational products. In the future, our embodied large models will enable versatile robotic platforms tailored to diverse applications, widely deployed across retail, logistics, industry, and home settings, delivering safe and intelligent solutions.