In the development trajectory of embodied intelligence, securing vast quantities of high-quality data remains a central challenge that the industry cannot sidestep.

If large language models rely on internet-scale corpora, then the development of embodied intelligence likewise requires large‑scale interactive experience. In practice, collecting such data is extremely costly. Hardware such as robotic arms is expensive to deploy—each unit can cost tens of thousands of yuan—and scaling up data collection is challenging. Moreover, data collection depends on highly skilled annotators and is time‑consuming.By contrast, in simulation environments, agents can engage in unlimited trial-and-error at a much lower cost and with greater efficiency, enabling them to rapidly amass vast amounts of interaction experience.

For this very reason, over the past few years, simulators have become a crucial enabling tool for the advancement of embodied intelligence, giving rise to a number of outstanding open-source and commercial platforms. These platforms have accelerated progress in robotics, reinforcement learning, and agent-based research, laying the groundwork for the field.

However, as research progresses, the industry is placing increasingly stringent demands on data: higher physical accuracy to ensure close alignment with the real world; richer interaction types that encompass complex scenarios such as rigid bodies, soft bodies, and fluids; and greater scalability and stability—capable of supporting both the microscopic dynamic details required in scientific research and the large-scale simulation needs of industrial applications.

Against this backdrop, we have launched a high-precision general-purpose physics simulation platform for embodied intelligence: RoboMirage.

“RoboMirage” has the following core features:

• A scalable contact modeling framework compatible with all contact types

It supports diverse contact interactions involving rigid bodies, 1D/2D/3D deformable objects, multi-joint structures, and various robotic end-effectors; features tight coupling simulation capabilities; is compatible with future differentiable simulation and high-precision training requirements; and allows users to customize and extend its functionality, providing a flexible, adaptable underlying architecture for a wide range of application scenarios.

• High-precision, highly realistic multi-body dynamics simulation capabilities

High‑precision, non‑penetrative, and time‑consistent contact‑force simulation supports strongly coupled dynamics for rigid bodies, soft bodies, and complex contacts—such as the tightly coupled interaction between intricate fabrics and robotic end effectors—while capturing microscopic dynamic details like static and kinetic friction and subtle force variations. Its accuracy far surpasses that of conventional position‑based dynamics simulators, making it particularly well suited to the complex simulation requirements of the robotics domain.

• Industrial-grade stability algorithms ensure 

By leveraging rigorous mathematical frameworks such as implicit integration and convex optimization, we accurately solve problems in continuum mechanics, ensuring algorithmic stability and temporal consistency throughout the simulation. This approach captures every dynamic detail and fully resolves penetration artifacts, meeting the stringent reliability requirements of industrial‑grade tasks like assembly and grasping, thereby providing sustained, stable performance even in complex scenarios.

• Pythonic design, with a simple and easy-to-use framework architecture

Prioritizing user experience, it features a developer-friendly interface that’s easy to learn, enabling developers to quickly integrate and customize the solution, thereby accelerating efficient simulation development.

• Advanced GPU‑driven heterogeneous acceleration technology

Deep GPU optimization fully leverages the GPU’s massive parallel computing capabilities, combined with data‑oriented programming, to deliver high‑performance, rapid simulations at industrial‑grade accuracy—significantly outperforming traditional finite element analysis and existing robot simulation platforms.

To more vividly demonstrate the powerful capabilities of “RoboMirage,” let’s first take a look at several classic magic tricks:

A metal ring hangs from the end of a thin string; When you release your finger, and it falls freely yet hovers steadily in midair. Two rubber bands, each looped over a finger of each hand, cross over one another and, after rubbing against each other, slip through with a gentle pull. A deck of cards is split into two piles, and with a flick of the fingers, they cascade evenly from both sides, falling toward the center… These seemingly magical moments, in fact, embody the most subtle principles of force and equilibrium that govern the physical world.

It is RoboScience’s physics‑based simulation platform, “RoboMirage,” that recreates these intricate and ingenious Magic Moments from the world of magic with high‑precision simulation technology. Leveraging precise computational models and sub‑millimeter‑level control capabilities, it transforms subtle real‑world interactions into computable physical processes, thereby pushing the boundaries of simulation technology once again.

First, let’s examine the classic magic trick, the Tomorrow Ring: “RoboMirage” can simulate the intricate contact and entanglement between a metal ring and a flexible rope, involving frictional sliding and rigid–flexible body coupling. This demands that the engine reliably handle dynamic contact, preventing interpenetration or simulation failures.

The primary challenge in simulating the rubber-band‑through‑ring magic trick lies in the mutual entanglement, stretching, and deformation of two elastic bodies, requiring precise modeling of their viscous damping, tension variations, and self‑collision dynamics.

The core challenge in simulating card shuffling lies in accurately modeling the contact forces and frictional interactions that arise when multiple cards are interleaved at extremely small time steps; the key is to maintain contact continuity and prevent interpenetration.

As for the simulation of the tablecloth‑pulling illusion, it requires high‑precision capture of the fabric’s instantaneous sliding, as well as the frictional inertia and force response of the object at the moment the cloth is removed, while simultaneously accounting for the non‑equilibrium dynamics of a rapid pull and the stability of the objects above.

It is important to emphasize that, while “RoboMirage” serves as the core infrastructure for bridging the sim-to-real gap, it does not constitute the entirety of our data ecosystem.

In addition to the large volume of simulation-generated training and validation samples, our R&D and validation pipeline integrates multimodal sensor data and operational logs—such as force, pose, tactile information, and video—from internet‑based corpora and knowledge bases, structured technical documentation, and user manuals (product manuals, CAD files, specification sheets, etc.), along with a small amount of real‑robot experimental data. These real-world and text‑based datasets serve as validation benchmarks for simulation results, enabling fine-tuning of perception and policy models, strengthening semantic constraints, and establishing baseline tests in realistic scenarios.

In other words, We possess both a high-precision simulation “foundation” and a research-and-validation system supported by multi-source data. — The synergistic interaction between the two effectively narrows the sim-to-real gap, accelerating the robust transfer and large-scale deployment of algorithms in real-world environments.

Furthermore, we have now completed the most complex, highest-precision, and most multi-step embodied manipulation task to date: furniture assembly.

Once the model has read the instruction manual, it can begin assembly: it deeply understands the structural logic of the components, enabling the detection, sensing, insertion, removal, and rotational coordination of multiple parts. It can also autonomously decompose multi-step tasks and execute coordinated bimanual motion across multiple joints.

Leveraging adaptive mating‑path planning and fine‑grained contact‑force control, the system successfully achieves a highly precise and stable assembly process, demonstrating exceptional performance in both component positioning and micro‑motion control. By continuously sensing the feedback forces generated during mating, the model can also dynamically adapt its operational strategy.

Even if the model is disrupted by user‑initiated disassembly during the assembly process, it can automatically restore its state and resume the remaining steps. With this framework, the system can analyze physical interactions across diverse scenarios—such as force feedback during robotic object grasping, deformation prediction, or motion planning—using a standardized approach, eliminating the need to develop bespoke algorithms for each specific object or robot.

By synergizing a high-precision simulation engine with a multi-source data framework, we have not only achieved unprecedented stability and accuracy in highly complex embodied manipulation tasks—such as fully automated furniture assembly—but have also laid a robust technological foundation for a broader range of real-world applications.

Going forward, we will continue to push the boundaries of integrated simulation accuracy, generalization capability, and real-world interaction, enabling robots to autonomously perceive, reason, and execute tasks in increasingly diverse and open-ended environments.

We believe this technological roadmap will accelerate the transition of embodied intelligence from the laboratory to the real world, giving rise to entirely new human–robot collaboration paradigms—enabling intelligent robots to become trusted partners and assistants in both everyday life and industry.

Acknowledgments: We extend our heartfelt gratitude to the open-source communities and academic institutions in the fields of computer graphics and robotics, including but not limited to Lin‘s Lab, HKU CGVU Lab, as well as outstanding open-source projects such as OpenUSD, softmac, libuipc, and diffclothai. We sincerely thank every pioneer driving progress in the embodied robotics community.