Visics

General-Purpose Embodied Foundation Model

Visics is a general-purpose embodied foundation model built on a VLOA (Vision-Language-Object-Action) architecture that is object-centric and integrates an embodied world model with a generalist manipulation model. It achieves generalization across three key dimensions: embodiment, object, and task. The upper-layer embodied world model handles cognition and simulates physical trajectories, while the lower-layer generalist manipulation model converts those trajectories into physical control signals for different robots. This enables a clean decoupling of high-level semantics from low-level physical laws, with pretraining on massive internet video and simulation data.

Embodied World Model

Unlike current world models that center on video prediction, RoboScience’s embodied world model is object-interaction-centric and 3D dynamic, with an architecture closer to the underlying representations required for real robot execution. Its core is not a video generation module attached to a VLA model. It does not need to render a grand world or operate at the pixel level. Instead, it builds the world model around object-level states, 3D trajectories, contact relationships, and physical causal changes required for embodied manipulation, forming a unified closed loop with the execution layer (the generalist manipulation model).

The model is pretrained on massive internet video data. Through a fully automated data annotation and cleaning pipeline, the team has accumulated millions of hours of object-centric, high-dimensional multimodal manipulation-related datasets (tens of millions of video clips), growing at hundreds of thousands of hours per week. The goal is to build a world-leading dataset of tens of millions of hours by 2026.

Generalist Manipulation Model

The model addresses generalization and dexterous manipulation challenges through an efficient closed loop of “physics engine – simulation data – end-to-end training.” It supports manipulation tasks across all spatial object types, including rigid bodies, articulated bodies, and 1D/2D/3D deformable bodies. It supports cross-embodiment deployment across different types of robots and end-effectors. It supports closed-loop manipulation and multimodal perception data including vision, touch, and force.

As the foundation for the generalist manipulation model to learn physical laws, we have accumulated tens of billions of high-quality manipulation trajectory data through our self-developed multimodal physics engine, RoboMirage. The 2026 goal is to build over 1 trillion high-quality manipulation trajectories (1 trillion manipulation trajectories across all spatial objects).

Breakthrough: Complex Furniture Assembly with VLOA

Based on the VLOA foundation model, RoboScience completed the world’s most complex, highest-precision, and most step-intensive embodied manipulation task in May 2025.

Furniture assembly includes nearly all the challenges in the field: in-hand manipulation, external dexterity, dual-arm coordination, precise force control, object rearrangement, and part insertion. Furniture assembly also demands extremely high precision, reaching the millimeter level.

The model can start assembly after reading the instructions: it deeply understands the structural logic of parts, enables detection, perception, insertion/removal, and rotational coordination of multiple parts, and autonomously decomposes multi-step tasks to complete multi-joint dual-arm coordinated motion.

With adaptive insertion path planning and fine contact force regulation strategies, the system successfully achieves high-precision and high-stability assembly, demonstrating outstanding capabilities in both part positioning and micro-motion control. By acquiring real-time feedback forces during insertion, the model can also dynamically adjust manipulation strategies.

Even if the assembly process is disrupted by user disassembly, the model can automatically recover its state and continue the remaining assembly steps. Through this framework, the system can use a standardized method to analyze physical contacts in different scenarios—such as force feedback, deformation prediction, or motion planning during robot grasping—without developing separate algorithms for each object or robot.